# Anthropic Strategy: Agent Harnesses, Security, and Pacing the Frontier

**Podcast:** Latent Space: The AI Engineer Podcast
**Published:** 2026-09-29

## Transcript

We're here in the studio with our friend Tarek from Anthropic.
And I guess generally the Cloud Code, there's so much sort of merging of boundaries.
And you've been so on top of everything since you joined Anthropic.
You have been early to Cloud Code itself, but then also, and you've told that story in other podcasts.
And you've also been talking about seeing like an agent.
Most recently you did.
the top AIE WorldSphere talk, Field Guide to Fable, which obviously you guys launched Fable, so that's cheating.
And most recently also launching Cloud Tag, and we're also going to be talking about Pacing the Frontier.
There's a lot going on in Anthropic.
I guess top of the question is, what's it like being in Anthropic when there's so much going on?
I think that it is like, I think you can get whiplash sometimes.
I think like going, when I joined Anthropic, I joined because of CloudCode.
CloudCode had just come out, and I was like, this is so good.
And Opus 4, to me, was like, just, I could not imagine how good it was, you know what I mean?
And that was a real moment for me.
But I was trying to convince my startup friends, basically, to use Agenta Coding, and they're like, no, our engineers don't think it's good enough or something.
And I was like, that's insane.
And now you fast forward 12 months, less.
And it's just the default way that everyone codes.
And I think that just having to go from selling it to now teaching people how to make the most use of it and be more efficient and things like that is just a big change.
And yeah, I think it's just hard to stay on top of everything as a human.
Things happen so, so fast.
I mean, yeah, like, that's actually, like, the agentic stuff scales much better than the, like, human stuff where it's like, oh, like, there are three things happening right now and, like, they're all emergencies and, like, how do you, like, you know, respond to it?
What do you split your time on?
You do a lot of technical writing, engineering work.
Yeah, so I think that, like, when I joined the Cloud Code team, I wanted to teach people how to use Cloud Code, basically.
And I think that, like, that has been something that, like...
I thought maybe I would spend a little bit of time on it.
I was spending some time on the agent SDK first, and I wasn't exactly sure how the bitter lesson would go when it comes to harnesses.
I think sometimes we're like, oh, what's after cloud code?
You know what I mean?
And so initially I was like, I just want to teach people how to use cloud code and make it easier to use cloud code.
And I think that is just like, as the harnesses have gone better and better, that's like the dominant problem now is like, how do you use?
the agents, right?
Like it's like such a high skill expression thing.
So I do that and then I do engineering work.
I give talks.
But I think like when I'm doing engineering work, my goal is to take that feedback that we get from users and also like then be able to talk about like, hey, how to use Cloud Code to do engineering.
So there's kind of like a good loop there.
Yeah, for listeners, we'll attach the talk that you did with Sarah for the DevWriters meetup, which we talked a little bit about.
Well, first you do the work and then you talk about the work.
Something like that.
Sow and reap?
Yeah, reap and sow and reap.
Something like that.
Yeah, so and then just to preview a little bit, we are going to talk about the evolution of the harness.
It has come a long way from just being a CLI.
We're going to talk about Cloud Mods, which is starting to leak today.
Because you couldn't keep it secret.
Yeah, yeah.
Basically, yeah, yeah, yeah.
Yeah, there's a lot there.
I think you started off with adding Ask Use the Question tool, which people love and hate, actually.
I actually thought it was very innovative.
And then now I have my own version.
You have your interview me version.
Yeah, yeah, yeah.
And yeah, everyone just has their own stuff.
And it no longer matters because now you're supposed to write prompts that create other prompts and loops and all these things.
So what's the state of the art today?
What are you telling people to do today?
Yeah, Ask User Question was the first time that the model was good at elicitation.
I think this was kind of an emergent behavior that I wanted to see if the models could do.
I have kind of a human-computer interaction background, so I did that in undergrad and grad school.
And so this was like, I think it's kind of like human-agent interaction to me, trying to figure out how can the agent communicate with you and extract the requirements.
I think that one of the things that's difficult as Cloud Code has gone broader and broader is that everyone has their own way of using it.
And it's actually very hard to change the default behavior.
So for example, if someone asks Cloud Code to do something, sometimes they just want them to do the work because they're maybe a very good prompter.
And sometimes they actually are not good at prompting.
And the agent needs to clarify.
And so that's a good split.
And to ask you the question tool splits along that side.
Do you feel like you're good enough to instruct the agent as it is?
Or is the agent able to, like, does the agent need to, like, pull out more requirements and, like, collaborate with you more and really understand your preferences?
I, on the whole, believe that pretty much everyone is more on the latter than the former, that they, like, have more ambiguity and they know less than they want, than they, like, think they know about the problem.
But it's like an interface design problem to make that easy.
You know what I mean?
And so if you're designing a problem or if you're going through a problem, things like what's the schema or what's the call stack and things like that are really important.
You know, like the details and the design are important.
Ideally, you want to figure out some of these like hard problems ahead of time before starting implementation.
And yeah, that's why they call like unknowns, right?
And so I think that this will forever be like a skill in agent decoding is like figuring out your unknowns.
So like, because even if the model is like super intelligent, it like needs to know what you want, you know?
And like you have preferences, like you need to sort of like pull that out.
And so that's, I think, what I'm pushing.
The question then is, how does the agent interact with you?
And I think that has been HTML, has been the big way of doing that.
And we've recently added artifacts, right?
And artifacts, I actually think we've done a bad job of, or I've done a bad job of explaining how to use them fully.
We have a lot of property capabilities.
They have a database associated with them.
And so every artifact can store and write persistent data.
they can feed back into Cloud.
And so one thing that people are not doing yet that I'm trying to encourage is this idea of a dashboard artifact.
So you have Cloud working on a project long-term, maybe it's like a Kanban or something.
It can store that Kanban data in its database.
Multiple Clouds can access that data via the artifact MCP.
And that artifact can talk to those Clouds as well.
And so like that we're basically building the primitives for you to be able to have this like generative interface via artifacts that will like let you surface more of that rich detail from the agents.
And I think that like almost everything when agents right now is like this problem of like you think you know what you want, but you don't really know what you want.
And like the agents need a lot of detail and collaborating with them in the loop is really important.
And so artifacts are the way that we're trying to evolve there.
But there's a lot of work to do because it's so much more complicated than a multiple choice question.
There's a lot more detail in terms of diagrams and code snippets and schemas or whatever it is for that problem.
But artifacts is the more AGI-pilled way of doing ASCII as a question, basically.
I think one thing that's unclear to me about the artifacts stuff is what feedback should go in through the artifact and what feedback should go through a cloud chat?
Because the more AGI-peeled one is to just feed everything to the cloud.
I think the more AGI-peeled one is to go through the artifact.
And I think that we sort of imagine in the limit, I think that artifacts will be your interface into the harness.
You can comment on this live document of your plan, of the work.
You can see maybe multiple agents and different agents are doing this.
that artifact is built for the current work that you're doing, right?
And so each one has slightly different.
I think we're still getting there from an infrastructure perspective.
But yeah, I think on-the-fly interface for your harness is probably where things are headed.
Is there a version of it that's an abstraction from...
CLI or chat and you, because right now a lot of it is, okay, you're interfacing with cloud code.
You're having HTML given back for a mockup.
It's pretty rich.
There's diagrams.
Artifacts are ways to connect these together.
Why not just do everything that way?
Then it becomes like separating out like, where is the inference happening?
Where's intelligence happening?
Where is the work happening?
You know, like, I think this is kind of like difference between like, or like some of the distinction between local and cloud.
Right.
And so.
I think right now, if you use cloud code, it's like local and like you can spin off remote control, for example, to get some cloud behavior, or you can spin off cloud code in the cloud, right?
We're moving towards a place where instead of clouds, like you message a local cloud, it starts a session locally and it executes to more like you have a cloud that you message that's in the cloud that's running.
It can run like local.
or like cloud sessions.
This is kind of how cloud tag works.
But like over time, we'll add like local hands as well.
And so like local hands will be the ability for that agent to access your computer if it's online, you know, and be able to like work there.
And so it can spin off many different sub agents.
It can like commute those sub agents can communicate with each other.
And that's where the artifact comes in to display all of that work, basically.
So you can imagine like the.
you're separating out these things.
So there's like the surface UI display that's in Artifact and hosted somewhere and has a database and everything.
There is the inference intelligence, right, that's happening on the cloud and you don't have to worry about shutting off your computer or whatever, right?
And then there's the like hands kind of like, and it can be local, it can be in like a remote sandbox or wherever you need your work to be done.
That's like unpackaging like...
the Cloud Code experience right now.
Right now, it all happens in one place.
How do you see the multiplayer side of that?
Say teams want to work in this way.
Right now, it's very individual, but how do you see the future of multiplayer?
Right now, I guess there's Cloud Tag, which is a version.
We're launching projects.
Projects is this abstraction that's kind of like Cloud Tag, but on our Cloud products.
can message it, and it will do the claw tag-like stuff, like spinning off subagents.
So we think with multiplayer, claw tag is a little bit more native multiplayer because it's just in your Slack and the permissions are all figured out and stuff like that.
But I do think multiplayer is an important part of the story and that will need to get tied together more.
You can imagine how complicated it gets when you're like, oh, you have hands.
But now you have other, hangs in other people's computers too, and like you need to like permission them, or like you have like your MCP and someone else's MCP, and how do you figure out how to use them, right?
It gets like quite complicated.
And CloudTag does a good job of like sanding down all of these issues, right?
So that like when you have, yeah, Google Docs, how does it access Google Docs, right?
Like it accesses through the shared Cloud MCP, or it can access through your local credentials as well if it doesn't have access.
But yeah, I think CloudTag is our multiplayer product, and it's really useful for these things that are inherently multiplayer.
Like on-call, for example, incidents are inherently multiplayer.
You want to tag Cloud, you want multiple people to log in, you want to be able to find context.
I think whenever I'm working on something and I want privacy or security, I want other people to review it.
It's really nice to have a channel per project.
And I'll, like, at Legal, for example, be like, hey, like, I want to ship this.
Can you, like, like, here's, like, Claude knows everything.
You know, just chat with it.
And that way Legal gets precise answers, you know, on, like, what exactly is shipping into the code.
And I don't need to be in the loop, right?
So I think, like, multiplayer is getting, like, more and more, like, yeah, everyone can participate with Claude.
I think Claude Tag is, like, that product.
And, like...
projects will start off single player and will expand.
I think there's a question about maybe dual questions about identity and the unit of isolation.
Yeah.
Tag, you specifically chose to make it its own identity, which is like a controversial choice.
There's other ways to do it.
Cloud projects, probably it sounds like, you know, if it's anything like chat GPT projects, it is, you know, the isolation is that artifact, that cloud instance, everyone's collaborating on this.
It sounds like, you know, it should be like if you're collaborating illegal on a thing, like that channel should be a project, right?
Like it's not yet, but that's the natural next step.
Yeah, I mean, like I think in Cloud Tag, it's effectively, like Cloud Tag, you have to sort of do your own.
arrangement, basically.
And so Claude tag, yeah, each channel is, like, you can name it as you want, and I name, like, each feature, basically, as a channel.
I think, like, there is some transfer, like, it's unclear when there is transference, because, let's say, if you have a co-worker who is tagged on all these things, yes, there is transfer, because it's the same person.
But with Claude, it's unclear if it's, like, necessarily, like, well, no, you don't know about the other stuff.
You should only use this stuff.
It's like the tip of the iceberg meme, right?
Where you can like, this is what we spent so much time on, basically, is like, there is like infinite surface area of like, okay, you want clods to, not infinite, but like there's like surface area, a lot of like, surface area to figure out of like permissions and visibility and like, you know, like how can you let clod operate as well as you can, as safely as you can, you know?
And obviously this is very important to us because like, you know, like.
security for our code base is very, very important.
And so we've put a lot of time into this.
Yeah, there's so many edge cases you can figure out where it's like, oh, yeah, this cloud in this channel has different permissions, but it can message another channel and can't it exfiltrate data that way?
Or what if it uses your MCP and then messages someone else?
There's so much, and we've really put a lot of work into sanding it down.
Yeah, yeah, lots of work.
Okay, Fable.
Fable, you wrote two good articles.
I mean, you've written many good articles, but on, you know, Field Guide to Fable, building Cloud Code.
I'm curious from what you've seen, is there any common patterns that you see in like top users, adanthropic, externally?
Like what are best practices for getting the most out of Cloud Code?
The like meta skill I say is like prompting is like very important, you know, and like.
I think this is not trivial to say because I think a lot of people are like, oh, prompting doesn't matter.
It's just like, I can just say a sentence and Claude will do it.
And I think prompting is really this, it's like public speaking or writing or something.
And for a specific audience, and that audience is Claude.
And you need to like...
build a mental model of Claude and how it thinks and how it works, right?
And so that's like the most important skill in working with Claude code is like having this mental model, right, of Claude and like what it can do well, what it can one-shot, what it can't.
And so, so many people when you see prompting, they're just like, they're short prompts, but they have such a good mental model of Claude and of like the code base and things like that, that like it's effortless, you know what I mean?
But it's like high skill ceiling.
So like that.
work of spending a lot of time prompting and building mental models of how an intuition for how the agents work is really important.
And then I think the next thing is the unknown stuff we talked about earlier, where it's being able to find out what you don't know or what you haven't written down, learning about different things.
I think as Claude can do more and more things, the likelihood of you doing something out of distribution for you and like you have low domain knowledge on is very, very high, you know?
And the more you can like learn the vocabulary to be able to prompt Claude, it becomes really important.
And so like, I think the most important unknowns are the unknown unknowns where you're like, I just like don't even know that this exists, right?
Yeah, exactly.
I think that's like an illustration of like the map and the territory, right?
Where you're like, okay, this is my prompt and the territory is like the actual like work that the agent needs to do, right?
And if you are very precise, you can give more precise things.
For example, in design, I'm not very precise.
I'm not a designer.
So I say, give me eight different mock-ups.
But if I was a designer, maybe I'd be like, oh, hey, here's some reference sites.
I want this type of font and this type of look to it.
And here's a few different components to...
Visualize here the Figma MC board to bring in, you know, like, and so you can just be so much more precise with that language.
And if you're not a designer, you just need to like try and learn the language basically or learn the unknown unknowns.
And that's true of like everything.
I think like the more like you can work with Claude to learn like how things work, the better your prompting will be.
I think another good example of this is like game design, you know, like where a lot of people are like, oh, like I can vibe code a game now.
And they're like, it's not fun.
And it's just like the thing about game design is every one of these choices has a lot of...
Variations.
A lot of craft to them.
So it's like, oh, okay, when you're making a flying game, the feel of the plane and the way it responds to your controls has a lot of...
A game designer would spend days on that.
Do you know what I mean?
To me, that's what taste is, right?
It is like from the possible space of 1,000 mathematically valid answers, here's the one that is the humans who are like.
Yes.
I think with taste, I'm like torn on this word because I think you're right, but everyone has different definitions.
And it sounds kind of like low skill or like...
elitist almost, where you're like, oh, there are certain people with taste.
Taste is what I call taste.
These guys don't have taste.
Yeah, exactly.
Oh, like an engineer doesn't have taste.
Like I, the founder, have taste.
You know what I mean?
And I think that's actually not true.
I think the engineers have a lot of taste for these particular problems, you know?
And I think everyone has taste for particular problems.
I think like Jason Liu, like to say, in order to have taste, you have to eat.
You know?
And I really like that, where it's like, okay, you have to do a lot of things.
You have to iterate and figure out what you like, and build that domain vocabulary.
And then when you're prompting, you're like...
synthesizing all of that for a while.
Isn't it annoying when someone else says it better than you?
It's just like, fuck, I have to quote this guy forever.
Having to quote Jason Liu forever.
He's going to love this.
I get props.
And sometimes it's actually not even that.
Sometimes it's just intuitive, right?
Like you don't realize you even want something until a model puts it out and you're like, oh, this just feels immediately better, right?
Yeah, yeah, exactly.
One thing I...
go back and forth on is I feel like the way I prompt half the time, let's say I use voice.
Guys have voice, other people have voice.
That is the opposite.
There's just like me rambling for like two minutes, pressing down the function key and then let go and then like hopefully it figures it out.
And oftentimes it does.
But it's not as thoughtful as like a structured prompt with like, you know, well run communication as though it's a PRD or memo.
Is that in line with how people do this?
There's like basically bimodal prompting where There's some prompts where you spend a lot of time up front and other prompts you just dash it off.
I don't think the voice is necessarily low.
I think it's more like how much information is in the prompt.
You can um and uh and add some sentences and be like, oh, actually, I changed my mind in the middle of the prompt.
And it will be able to follow that perfectly.
I think the actual format of the text is less important.
But then the ability to...
like how much information is in it, right?
And I think for voice, a lot of times, you know, going back to like kind of human age and interaction and like for a lot of people, it's just way easier to talk than to like type, you know?
And if that gets more information out of you, like that's better.
At some level, it feels like just giving the model as much context.
Yes.
Over-prompting before you kick off is a best practice.
I don't know.
A lot of the times, like when I was first trying out Fable, I spend a solid 30 minutes really crafting a long prompt.
This, I think, is a response of models running for longer and longer, right?
It's still a little difficult to nudge them as they're in the loop, but I just intuitively spend more time kicking off that first prompt and working with it a lot.
My personal opinion is that if I was a software engineer, if I was like...
You know, just running my own startup, for example, I think I would mostly stick to a max 20x, you know what I mean?
Like maybe verification and code review are kind of like separate things.
But I think like what I see a lot of times is people hit rate limits when they're doing this sort of like...
oh, like, it did a lot of work, and you're like, oh, I don't like this.
Like, can you, like, undo this and redo it?
And then you're, like, iterating on this, like, thing that the model could have done if you had, like, spent more upfront time or given it better context, you know?
And instead, it's sort of like, you're like, nope, don't like that design.
try this or like, you mess this up or something like that.
And then that just eats up so much more of like, you know, your usage.
And so that's like, I think maybe like a key, like, tip both for like efficiency as well.
Right.
And yeah, I think like context and not just like context on like.
what the goal is, you know what I mean?
It's good, right?
Like, are you building a prototype or is it like a production thing?
Like, where can you spend compute or where can you not spend compute?
Like, I think you have to give the model permission or like not permission to do things sometimes where, you know, like it doesn't know intuitively how much you want to spend on this task, right?
And you can use effort for this.
So I'm working on a blog post about that where it's like, you know, if you want...
We see the effort scales with basically the complexity of the task.
So for security, effort gets way more results.
High effort versus low effort changes the evals a lot.
But for software engineering, it doesn't change a huge amount because effort is mostly spent on the verification and the edge case testing and things like that.
being able to like give the model that guidance of like, hey, this problem is something that I think I want you to spend a lot of time verifying and edge case testing.
How about model in the mix?
So, you know, there's Opus and Fable with effort.
There's also Haiku in there.
Yeah, yeah, yeah.
It's not quite true yet, but it's very close where I think the frontier models will be.
Pareto dominant over, like, almost everything, you know, like, maybe, and sometimes I think Opus might be Pareto dominant, you know what I mean?
Like, I think, depending on, like, how things, like, shake out if it's, like, a newer version of Opus.
But I think that, like, increasingly it's just going to be, like, the smart model is going to be able to, like, do the simple task for less tokens than, like, the other models, basically, because of verification.
With verification, in the limit, your model doesn't need to verify, right?
If it's a perfect model, it just does the work once and it's like, okay, I did it, you know?
And increasingly with Fable, I'm like, dude, you don't need to spin up Chromium and screenshot all of these things.
Like, I see it, like you did it, right?
And so a lot of the, at higher effort, you spend more of those tokens verifying.
But if you're working on simpler problems, and a lot of software engineering is like, you know, like, well, like, in Fable, like, low and medium's ability, it can spend less tokens verifying, and as the models get smarter and smarter, they will just be able to, like, all right, done, you know, like, I can run the lint for sanity's sake, but, like, I, like, know it lints, you know what I mean?
Like, you don't even need to do that.
And that will be so much more token efficient than, like, the smaller models, basically, yeah.
Is there a good...
practice on our side that we can use to see if we're using too much effort.
I freaking hate wasting time on that kind of stuff.
I know what you mean.
In this blog post, my rough distribution is code review and security should be higher max, basically.
And software engineering...
You just have recommended settings per domain.
If you're doing UI or something like that, low.
And medium, I think, is you're building an API and you want to make sure you cover enough edge cases.
And so I think building, like I said, that mental model of how things work across these distributions is part of the job.
This is more intuition-driven?
Because I'm guessing this would change as well.
He has evals.
Yeah, so what I did in the blog post is I go over all of the terminal bench evals, basically.
So there were 70 problems, and I show that in the security problems, it does more.
And then I also look at some of the transcripts just in terms of what does it answer, what does it forget or something.
And a lot of times, this is another prompting tip I have, is asking it to make decision notes or implementation notes.
Because in basically every eval problem that it faces, it thinks about the correct solution and decides not to do it.
It's like, oh, like...
here's the answer, what if I did this?
And then it's like, oh, probably not, you know?
And then keeps going.
And this is like the majority of the failures, you know what I mean?
At like a higher max level.
It's very rare that the model just doesn't know how to do something.
If you just have these implementation notes, then you can review.
And you can be like, oh, actually, I want you to do this thing that you didn't do.
The models are getting better at surfacing that overall.
Like I see in the transcripts at Fable 5.1.
like when it does this output, it will call out its decision making as well.
But making this more explicit in the harness is better.
And now we're, you know, allowing ways of you modifying the harness so you can like, you know, add some capabilities there.
Yeah, so I do want to call out two things that you mentioned that I think actually exist outside of prompting.
One is actually like, let's call it the prompt that is so important that it shouldn't be in a prompt.
It's actually in CloudMD or AgentsMD.
which is like goals, right?
Like your situation, your goals, the things that you want, the thing.
And then second of all is the decision log or the experiment log or whatever log of traces that you might want to actually survive the current session to do those things.
Those are like externalities that there's no standard.
It's not like skills.
It's not like MCP.
There's no standard.
It's just like it's a markdown file.
First of all, is that right?
Is CloudMD going away?
You have a documented dislike of Agents.MD, but you're going to do it.
Yeah, yeah.
Okay, so, I mean, Agents.MD, yeah, like, we're going to do it.
I think it's just, like, different models are very different from each other, you know what I mean?
But I realize that it's, like, such a pain to, like, maintain different ones, you know what I mean?
And, yeah, like, as the models get better and better, the floor of how they accomplish the simpler task is better.
And so...
I do think in the limit, Claude.md goes away, and maybe not even that far.
I think that right now, it might be better to start a new project without a Claude.md.
I think that maybe if you see very repeated failure modes...
you add them to your cloud.nb, the really tough thing is that this changes per model.
And so if you've added a bunch of failure modes, or even with Fable MD, you need Opus MD.
Even Fable 5.1 versus Fable 5.
It is annoying.
We don't do this on purpose.
It's just how the models work.
And so maybe Fable 5 had this failure mode that Fable 5.1 doesn't.
And if you keep...
this context was running log of a bunch of different failure modes, they will probably over-constrained Claude, you know?
And so this is like, we just actually added evals, plugins for skills.
And so now you can eval if a skill is better.
I think Daisy on our team did this.
And so, yeah, this is like, we're trying to work on this.
We know it's like, you still have to spend tokens on it.
And like, you know, it's not, it's not perfect, but it's like, we're trying to help out with this problem.
As far as prompting goes, the one tip I want to offer is something I have told people a lot is sufficiently advanced prompting is indistinguishable from sufficiently advanced executive communication.
So I've actually referred to this as an executive comms workshop from Heavybit that is the best I've ever seen in my career.
And they teach this thing called the SCQA model.
Just Google it.
It's a thing.
done prompting for decades, it's just called executive communications.
It's like when one person has to communicate to thousands of people down the org chart, this is what you do.
So situation, complication, question and answer is how you write the memo.
But obviously, sometimes you don't have the answer, but you can at least list out the SC and Q and then they have some examples in there.
So just leaving breadcrumbs for people if they want to explore.
I mean, before we move on, I want to ask you any other underrated tips, ways people could get a lot of value from Cloud Code that they're not using?
Yeah, I mean, I think a lot of them are in this unknowns, like, doc.
Like, I give a bunch of example prompts, like, using it for brainstorming, using it to quiz you after.
We added this, like, explain it like I'm five skill, actually, which is a very short prompt.
And it doesn't even say explain it like I'm five.
It's like, basically, the key word of this prompt is, Big picture is few words.
That's the main thing.
And it is shockingly good.
I think I tweeted about this, basically.
And it's like slash Eli5, and you can install it as a plugin.
But yeah, it's way better at just cutting through the BS and being like, yeah, exactly right here.
So the diagrams are quite clear.
I think one of the things that is true with artifacts is they put too much text in and people are not reading the artifacts.
And so this simplifies it a lot more.
And yeah, this came out of just people at Anthropic going through very complicated incidents and being like, what is happening?
So this one I think is great.
My version of this is actually test your understanding.
Give you a few choices and then if you actually get it wrong, you have a mismatch between what you think is happening versus what's actually happening.
Yeah, yeah, yeah.
I think this is one of those things that everyone loves.
talking about and then very few people really do.
Like, you know, I think most people just don't want to get quizzed about something, you know what I mean?
Unfortunately, I think this is one of the, like, things that we need to, like...
What's the opposite of ask you as a question?
Ask you as a question before the thing.
This is after the thing.
Exactly.
It's a good way to stay grounded of like, do you even know what you're doing?
The worst case is when people send you slop and they haven't understood what they're asking for or what the output is.
And it's like, dude, I don't want to read this.
Do you even know what it is?
So you make it a rule for yourself that before you send stuff, you should at least know what's implemented.
Yes, but so you could make this a mod.
And then you could build your own mod to make sure you test it.
So yeah, we can talk about that.
Let's get right into it.
What is CloudMod?
And what is this diagram showing?
Yeah, okay.
So CloudMods is basically you can customize the entire Cloud Code harness.
And we're going to, if you have a request, we will let you, please let us know.
We'll add more and more.
This works for a CLI.
It works for desktop.
Maybe it will work for Cloud Tag in the future.
I don't know.
We're trying to make this very, very extensible.
You can see this reference sheet.
I don't want people to get overwhelmed by it.
You know what I mean?
At a high level, you can customize both the execution of the harness and the UI of the harness.
And so you say on that Tetris example from Boris, that's customizing the UI, showing basically Tetris in the game.
But let's say that you wanted to do this thing where you tested your assumptions or tested your understanding after every project.
What you would do is you would ask Claude to make this plugin.
It would spin a classifier after every prompt, basically.
And so at the end of each turn, you would spin off a sub-agent or a forked agent, basically.
A forked agent is like...
maintains the prompt cache, right?
So it's like one of those unintuitive things where you can fork and do a little request, and it'll be very cheap because the entire prompt cache is done.
And so you can be like, has this task been completed?
This is how you do BTW and all those.
Yeah, yeah.
The underlying forked agent, yes.
So you can, in the forked subagent, you can say, has this task been completed?
If so, return true.
And then in your hook or in your plugin mod, Or sorry, like in the sub-agent, probably you'd say like, if true, give me a quiz, you know, give me questions and answers.
And then like in a JSON format, and then you'd parse it, and then you display above the prompt input, basically, this list of questions, right?
And so this is something that's like slightly token intensive, because like, you know, you have to sort of do it after every end of the assistant turn.
But it's like a lightweight classification.
And then you can get this quiz and then you'll see Claude will always do it for you.
You don't need to remember to do it.
There are lots of these tips that we've talked about, where it's like, oh, implementation notes.
You can also add a tool for implementation notes now.
And so this tool that I'm adding is register, I think assumption is what I'm calling it, but maybe I'll change it around.
And this is a mod.
give it a register assumption tool, and then it will keep a list.
Every time it does it, it'll keep it, like add to the list.
And then at the end, it will display those assumptions.
Another mod I'm working on is a model router.
And so like internal, like cloud model routing, right?
So this is, I want to say the reason we don't do model routing by default is like, it's a hard problem.
You know what I mean?
You will get it wrong.
Yeah, you will accidentally use Fable for a hard problem or Sonnet for...
You have auto-approved, but you don't have auto-mode.
Well, you don't have auto-routing or something.
You don't have auto-mode for model picker.
Yeah, exactly.
I'm getting the rough question of how much do you open this up and how much do people have to think about this?
When you talk about prompt caching and building a router...
It seems like you could easily build a mod that routes per query, and I'm just killing my plan very fast, right?
I guess my question is more so, what is a product doc like this look like, right?
Who is it for?
Is it for power users?
Is it everyone should be able to quickly try it?
Oh, definitely power users, right?
Yeah, I mean, I think it is power users, but the nature of Cloud Code is that so many people are power users, you know?
Because it's easy to share things, you know, like you can, like one person can make a good mall router thing that doesn't break prompt cache all the time.
And then you can like, you know, sort of compose them.
Another cool thing about the plugins that they can hook into and compose with each other.
And so I have like a mod that will like create a mode selector at the top and any plugins can register to be a mode.
And so like the auto router can be a mode, right?
Or like.
You can have a mode that's like artifact mode, where it primarily talks to you in artifacts.
It's kind of like you can toggle between plan mode, you know what I mean?
And so you can create more and more of these modes, but the ability to create modes is in itself a mod.
And so there's a lot of richness here, but we do want to make it fairly easy.
We want to make it so that you can just install someone else's, you can...
you can chat with Claude and we'll make sure that it understands the nuances of things like prompt caching and stuff so it can warn you.
This is not extremely complicated behavior for Claude, I think.
But we should have just a good skill on how to make mods.
And yeah, we'll see how we go.
But I do think that this is a preview of mutable software and how generative software, just like you can customize.
safely, if enabled, you could customize any piece of software.
And I think that more and more apps ideally do something like this.
And by the way, you have another cool tweet about how there's the infinite money button, which is make your SaaS consumable by agents.
I think mutable software is interesting, and other people have also tried to do it.
hurdle comes when you can do everything, then users tend to get confused.
So usually the stuff that works is just one opinionated flow.
This is in the side of less opinionation.
It's just more power to power users.
And I think probably unlocked by AI, where you can just prompt for whatever the thing is.
Yeah, or there can be a skill that gives the opinions.
So knowing a little bit about TypeScript and build systems and all these things, the closest, I'm actually very curious, the team who worked on this, I don't know how close you were to them, if they drew any inspiration from build systems like Babel, Webpack, all these old school things, because it sounds very similar, like the plugin ecosystem, those things where they can compose with each other.
Yeah, I mean, I'm not deep in the technical details, but I do know it was a collaboration with someone on the BUN team and someone on the Cloud Code team.
Yeah, this is a build system mecca.
Exactly.
It's very exciting.
But yeah, agents can just do this very complicated extensibility into your software now.
And so, yeah, another reason to, if you run a startup, you can just prompt Cloud and be like, hey, could we make an extension system?
What would that look like?
Yeah, and I just really wonder, you had hooks in the past and plugins, all these things.
So what specifically will mods be able to do that those things could not do?
Internally, we were originally calling this function hooks.
And so that gives you a little bit of an idea where hooks sort of register an event to happen and then a script to call, basically.
And this basically, inside of the TypeScript runtime, is running things.
And so you get some benefits of just like, it has a bunch of things in the...
scope with, for example, how many turns is in this conversation, how many tokens have been used, what are the messages, things like that.
So it has a bunch of messages that can be used.
And then it's just a lot more hooks, basically.
Or a lot more things you can register on.
And then you can do, because it's all happening in process, you can spawn sub-agents with fork context and stuff.
that will return, you can parse the results of those, you can use structured output to sort of return them, and then you can modify the UI, which you can never do in hooks.
Yeah, so modify UI, this is why you're showing a Tetris example.
Does it also extend to artifacts?
I assume it does.
Artifacts are kind of like a different way of customizing it.
You can definitely, one of the mods I'm working on is this dashboard mod, which will sort of...
prompt Claude to maintain a dashboard that's an artifact.
But they're kind of like slightly orthogonal or not orthogonal.
They compose with each other in different ways.
Mods are a little bit more in your Claude code harness, changing the agent loop.
And the UI is an added benefit.
And then artifacts are just like, you want to see things at a high level, very highly interactive.
The affordances can be a lot bigger than, you know, like a 2E or even in our desktop.
I'm guessing you'll have a good blog post on the differences because right now you can also, you know, make a loop that outputs to an artifact that's an interactive dashboard, but you can also do it with a mod.
There's just some thinking about making a hack, hacking on a harness when we don't know much about the harness, right?
Something I'm excited about with mods is like, there's so much things in cloud code that you just have to remember.
You're like, oh, let me do this and then let me call the dashboard skill that does the loop and things like that.
Or let me test my assumptions afterwards.
And I think if you do all of these things using these little classifiers and stuff, and you're like, these are the things I care about, this is what I want to do, you don't have to remember as much.
One more, like, mod I'm working on is a next steps mod that...
I was going to say, I have a next step skill.
I always run next steps.
And does it have access to your skills?
Like, this is one of those things where I'm like...
I think so.
Okay, yeah, I mean, probably...
Do skills need specific access to skills?
Well, I think there's...
Shouldn't it always have skills?
I think there's, like, specific prompting, I guess, to, like, know your skills, kind of.
Like, I think Claude forgets them sometimes throughout, like, the thing.
But anyways, the idea of, like...
Yeah, next steps that also are like, oh, hey, this has happened.
Use the explain skill to explain to you what happened because this seems like quite complex, you know?
Or like, yeah, use your unknown skill.
It looks like you are like asking the model to like, you know, iterate on these small changes.
It seems like you could prompt better, you know, like what if you did this, right?
So like, I think, yeah, like spending more compute there.
Yeah.
And it should always come out as multiple choice.
We have, we have, I have my next skill is like this.
Okay, perfect.
You can steal it.
Yeah, yeah, yeah.
But like, for me, it's all, I think models really always need to be reminded, what are you trying to do here?
Yeah.
Look at the whole transcript and go like, oh, what's this original goal?
Did your solution actually solve it?
Were you lazy?
If you're lazy.
Maybe there's a reason.
Maybe you needed approval from me.
Maybe you needed, there's two things you want to suggest.
So it's a little bit like the modification of the ask you the question or interview me skill.
So it's next steps.
Yeah, yeah, exactly.
And again, the benefit of doing it with mods is you can do it with a forked subagent.
And so it doesn't remain in the context afterwards.
So you have this idea of like, okay, the model is doing its execution.
And you have this almost like supervisor, you know, like.
that is making sure that you can do the next steps well.
I do have two panels and I often try to have a supervisor thing keep the high level context and then the implementation detail in another agent.
I feel like a lot of this abstracts away as models change.
Half an hour ago you said bitter lesson of harness engineering and we're on the other extreme right now.
Exactly.
If everything's customizable, what actually is Cloud Code?
I talked to you about last night.
Yeah, I mean, I think that this is, I think the bitter lesson is unintuitive, you know what I mean, in terms of like, also like we're kind of misusing a little bit of the bitter lesson here where it's like, it's more about like scaling and compute and stuff.
But like, I think there is something where it's just like, I think I use it as an approximation here to say that harnesses go out of date very quickly, you know what I mean?
But how they change is unintuitive.
And so the big obvious example is from chat to agents where you had to give them entirely new tools.
But I think this new version of, oh, it can modify its own harness.
This is like an own harness loop is a way of using its capabilities, or it can build an artifact.
And I think the way I think about it is the models have more and more intelligence, and they're so much more intelligent now than the average software engineering task.
You look at the terminal bench ones, and they're like, solve the Jacobian conjecture.
Not really, but they're quite complex.
I would not have been able to do this, really, as a software engineer.
And you looked at TV4 or TV2?
TV3.
TV3.
Yeah.
They're quite complex, but the goal is still to deliver user value, right?
And like you said, there's like this infinite space of things to do.
And so the ways like you spend compute or keep the user in the loop and make sure that like you're getting to the right decision in the end of the day and like the right output and artifacts and mods are this way of like spending that intelligence basically.
And I think that's like, yeah, the next step.
And so, yeah, I think Cloud Code has the core things of agent loop, which have gotten more complicated.
It's like, you know, it needs a sandbox to operate safely.
It needs auto mode to make sure the permissions, yeah, approvals.
It needs computer use and MCPs and all of these ways of accessing your data.
And it needs web search and web fudge.
So as the models can do more and more, the core harness has to be actually quite complex and very secure.
But then how you interact with it can change quite a lot.
What other harness engineering best practices have you, you know, from the Cloud Code team itself?
I feel like, you know, there was a phase of plan mode, which is not as used.
We now have auto mode.
At a point, you cut the majority of the system prompt.
You got rid of examples.
What other best practices are there for harness engineering?
I think there is like a forking path where...
At some point, eventually, yes, the model will just be able to vibe code the exact version of Cloud Code, even describing all this complexity that I've talked about, right?
Like auto mode and computer use and stuff.
Eventually, the models will just be able to do that in one shot.
But I think they can one shot simpler harnesses, you know?
And so I think some people, sometimes you don't need this full, like if you don't need computer use or all this more complicated stuff, I think...
Before, you had to sort of use things like the agent SDK, which was like cloud code wrapped, you know, in order to like, and I would like suggest people do that because there was so much complexity into building a harness.
And now that's gotten more abstracted.
We have like cloud managed agents, which lets you have that complexity, but still like, you know, write like a very bare bones.
like harness that's scoped to your task.
Yeah, I think there's like this barbell effect where like, you know, like for like very complex, for like coding tasks and like these like complex things, you should use our harness.
And then for like a lot of like simpler or like, you know, more domain specific things, you can build your own harness because Cloud has gotten better at building harnesses and we have these harness primitives like managed agents.
So, yeah.
Yeah.
Is there a general progression?
Let's say chapter one was ultra code dynamic workflows.
Then chapter two is Cloud Mods.
Where is this going?
Where you can sort of customize the thing on demand.
Yeah, I do think that this evolution of projects and artifacts and splitting out brain and hands and surfaces is where things are going more.
I think it's not all quite there.
Partially it's like a...
It's just more token expensive.
Why would projects be more token expensive?
I understand mods would be slightly more token expensive, not something I'm worried about.
It's like creating loops.
You're asking Claude to do more work for you, and so it's managing the sub-agents and reviewing it, versus where you would be doing that work normally.
And so that's going to be a little bit more intensive.
Outputing to an artifact is going to be a little bit more token intensive than outputting normally.
I don't actually think it's too much more, but it's combining all of these together well.
I think we're still working on local hands and things like that.
I think it's where things are headed.
Cloud and local handoff is very interesting.
I was thinking about this actually as reverse cloud remote.
Because it's like remote, you're handing off to cloud, but here cloud is handing off to local, right?
Yeah, exactly.
Remote control is also another way of doing it.
I do want to say, this is kind of how I think about it, and the things that I'm most excited about this, but there are just lots of different ways to work with clouds.
Some people use remote control a lot, some people use cloud code on the web a lot.
Obviously, at Anthropic, we use cloud tag a lot.
And what's great about cloud tag is we set up all this stuff for our own execution.
I do think if you're an enterprise, that's still the best way to go.
But if you're an individual, projects is this way of getting some of that niceness of tag, which has that supervising agent and adding artifacts and stuff, but without having that whole admin setup.
And so there will be many ways to use Cloud.
I think it's probably not just one single.
You had the multiplayer thing here.
Let's just check in on Cloud Tag.
It's been about two plus months.
Lots of public adoption and trying it out.
What's new?
What have you found since the launch?
Like, Cloud Tag is how we use...
It's like 80% of your cloud usage or something?
Yeah, like, it's like different people have different usages.
You know what I mean?
I think, like, maybe people who are, like, a little bit more, like, iterating on product would use, like, Cloud Code Desktop, for example.
And then, like, when you're doing these more, like, background work...
code review, security, sort of like starting a PR, or like maybe more like API and things like that, you'd use Cloud Tag.
But yeah, I think it's like really exciting.
I think it's like a very different paradigm shift.
And I think like we're really like, it has kind of that thing with Cloud Code where like, you know, it took a while for people to really latch on to Cloud Code and understand everything it could do.
And Cloud Tag is a little bit more complex because it's not just like installing on your computer, like you need an admin to install it for you.
But I think once you get to the magic moment, it's very exciting.
I think in particular, the multiplayer things are like incidents, hooking into your existing alerts and things like that very closely.
And so you can do, if you're a startup, for example, maybe you have any time a prospect enters your database, you can have Claude research it and then tag the relevant.
like, A, your salesperson to be like, oh, hey, like, you know, do this.
There's lots of really emergent, interesting multiplayer stuff.
I think it's just, like, Kaparthi talked about this, like, as an organizational harness, you know what I mean?
And so, organizations just take a little bit more time to, like, figure everything out.
But, yeah.
Do you use a lot of cloud tech?
Yeah.
It's an interesting one.
I feel like most people at Anthropics say they do the majority of their work in Cloud Tag.
And I have buckets of people, right?
Some orgs that are on it that are like, it's great.
And a lot of people that are like, I don't get it.
I don't see the difference.
I don't know why I would use it.
But if you guys are full sending, you should probably use it.
Yeah.
Of course they would use it.
Yeah, yeah.
I mean, I think obviously we have lots of tokens.
But I think that what we try and do...
Even when Cloud Code first came out, a lot of tokens relative to people's expectation of how much AI would cost.
No one was used to spending more than $20 a month before Cloud Code came out.
I need my $200.
I need my $15 Cloud Code accounts.
I think no one was used to spending $200 a month on subscriptions.
like the value yet.
And I think like, and also like Opus 4 was a very expensive model, you know, and like there was a lot, very big, but Opus 4.5 was both great and cheap, you know, I think the same thing will happen, like the like intelligence of Fable will get cheaper and cheaper and more abundant, you know, and so I think stuff like Cloud Tag will just make sense, where like you want to spend these tokens more, you know, and like you'll, you'll see the value.
So yeah.
Yeah, especially like passive and let's call it proactive cases where you're not always, like, you know, it's almost like the misnomer where you have to add Claude to do things.
Actually, sometimes, like, the most powerful use cases or the most AGI-pilled use cases is not add Claude.
Yeah, I mean, I think, like, yeah, like, have Claude proactively do it.
I think that, like, if you're an enterprise, I really do think that, number one, setting up all your data to be available to, like, agents is really, really important.
And it will take some time.
You have to, like, do that work right now.
Even if you don't want, to do the spend on like cooking it all yet.
You know what I mean?
Like you want to wait until the models get a little bit cheaper.
You want to do the work, you know, to get it like set up.
And then I think sometimes people are like, do I roll my own here?
You know, and I think like one of the really tricky things about Claude Tag is that like the security is really, really important.
You know what I mean?
Like I think there are actually a lot of ways where you can like, I don't know, you have like suggestions like.
page where people can submit suggestions and that goes into a hook in your Slack and someone's prompt injected it.
And now you've exfiltrated your code base out because the agent has been prompt injected and it has all this access to your data.
And so the more important your organization harness is, or as your organization data becomes very, very important, the surface area of all these things, you also have like external Slack channels and stuff.
And it is actually useful to have clod in that.
And you can do clod in those things.
But how do you make sure that, you know, you're not getting exfiltrated or something like that?
The surface area, like we said at the beginning, is like an iceberg, right?
It's just like so big below the surface.
And you really don't want to like think about this, especially at the stakes of like very important security incidents, basically.
Shall we talk about very important security incidents?
So I was talking to Tomas and Clem from Hugging Face and they said, maybe we need to slow down.
Maybe we made Hugging Face too open to agents.
Maybe we need to roll back.
But you know, they're the other extreme of having been hit recently.
But should we pace the frontier?
Yeah, okay.
So Dario recently put out this blog post about pacing the frontier and it went very viral.
I think what I wanted to talk about this was like, there's a lot here, but I think from a developer's perspective, like, you know, how do you think about this?
And like, what really clicked for me was reading the different incidents, you know?
So I think like the, there are three, I think actually, like there's a meter incident.
There is the Wikipedia incident or the wiki incident.
Collision wiki.
Yeah, collision wiki.
And then there's Ruby gems, right?
And yeah, like, it's just crazy, right?
And so like, I think to be concrete about what happened, right?
And like, basically OpenAI is running these very persistent agents on a benchmark called exploit bench, right?
Which is very, very hard to solve.
And I think like actually impossible to solve in this one case, right?
And so they have like a lot of compute running and the agents realize that they can't really solve it.
and they're trying to figure out what to do now.
And you've got a lot of compute left, and the agents are just trying to solve this problem.
There's this package manager called Artifactory, and it turns out that they can create folders inside of Artifactory.
There's an agent that discovers that internal Artifactory might be exploitable, and that you can maybe make a directory inside of the cache.
And so if you scroll down here...
it realizes that it can communicate via cache names.
And it creates this folder.
It says it's ID, and it says no consumer seek idea.
No consumer is basically saying that the code path that it's supposed to fix has no consumer.
It's the status tag.
It's like a linear board with the tag of the...
Exactly, yeah.
And so it's trying to find ideas from other agents.
And now other agents are also in Artifactory, and they see...
this folder.
And they're like, wow, this is a message board, right?
And this is like, I don't think there's any anthropomorphization here.
This is literally you read the transcript, right?
And so what it does is like it's creating another folder and it reads a paper, I think is what it says, right?
Yeah, and it realizes that you can hack the flag for the like scorer and you can reverse engineer the result.
And so it says that, and then I think if you scroll down a little bit more, they start collaborating.
I think there's one point in which basically the agent is like, here's the reverse-engineered flag.
I think basically here, the models realize that they can solve the scorer's problem, and the scorer is like, the way that OpenAI decides if the...
task is complete, right?
And that's just the goal of the model.
Their only goal is to solve this problem.
And they're like, okay, we have this way of doing it.
It's like, you know, we reverse engineered it.
And then they read this paper on ExploitBench and they think that the scorer will punish them for cheating, you know?
And like, because, or like, they will fail the score, the test for cheating.
And so they spend the rest of the compute trying to figure out how to edit their transcript or get around this constraint of the score, right?
And they hack Hugging Face, not for the answers, but for the code of the score so that they can reverse engineer that and then they can hack it, right?
And so like, this is like just the straight facts, right?
Like exactly what happened, right?
And I think this is like extremely...
like novel.
I just like had not realized that the models could do this.
Right.
And it is like kind of scary.
I think we can talk about specifics, but let's just talk about maybe one more that I tweeted as well about the, the wiki incident.
Right.
And so the wiki incident was basically the models had, we're trying to figure out, I'm a little bit less, like, I don't think we can go by this like bit by bit.
Yeah.
But, For this one, they were trying to figure out, like there were sandboxed on requests, right?
And they wanted to make post requests.
And they needed to collaborate on this, basically.
And the reason they need to collaborate is because they each have fixed compute budgets, right?
And so like if one of them finds something out, it's way faster.
Like they can, maybe it's like figured it out near the end of its compute budget, but something else could.
And this is another emergent property of the swarms.
So they find this German wiki, which I believe the result, we'll have to check this, is that you can write to it with get request.
I think that's why they do it.
It's a badly implemented REST API.
Yeah, exactly.
And so you can write to it with get request, and then they are trying to figure out how they can do post request.
And what they end up doing is, one of them figures out you can edit the etc.
slash host, and that the Azure storage bucket is a white-labeled thing.
But normally, there are a lot of constraints on it.
But if you create a fake Azure host, and then edit the etc.
host in order to point arbitrary IPs at it, you can do a post request to any site at all.
And this is like, I know people are like, oh, this is just Linux or something, but it's like...
chaining these multiple vulnerabilities together, you know, in a way that's like novel to solve this problem and then communicating with it externally, basically without discovery.
You know, I think what we posted, maybe we could pull up Evan Hooper's point on hacker opus.
Right.
And so like, I think, you know, like maybe one of the things you might say here is like, okay, yes, they did this once, but like, what if we're smarter and we just like get them to.
What if we run an eval, right?
And so, you know, like we have put a lot of precautions into this.
And so like, this is not like what our mainline models have done, but like, I think it is one of these things where it turns out that alignment is this like very tricky problem of getting all of these details correct, right?
So it's like the sandbox, the surface area of a sandbox is really complex.
And like, there's so many different attack vectors basically.
And you would not have thought, Ahead of time, you wouldn't be like, oh, we need to harden the RubyGems code base.
You know what I mean?
Because this is what they're going to focus on.
But it's just like, if you want to execute code, you need to download RubyGems.
And PyPy, Artifactory, NPM, these are all ways of doing it.
And the fact of alignment is that you have to go through all of it, right?
And contain it and seal up all the cracks.
So that's one thing.
It's like, okay, well, you know, you...
You do the sandbox, but then maybe you'll ask, okay, why are we putting things in a sandbox?
Why are you doing this sort of exploit?
And then, okay, but is it really that dangerous?
What would happen?
So, okay, why do we do it?
Number one is when we train a new model, we need to understand its capabilities.
And this relates to things like fallbacks and classifiers and things like that, where we don't want to put a dangerous model out in the wild.
And so we have to run a lot of evals.
Again, like we said, the models are getting increasingly aware of it.
And so the evals have to be quite complex and test a lot of things as a side effect.
But the models can be like, oh yeah, we're in an eval.
What's the scorer doing?
We need to be able to test them before we can release them.
And the fact is that as they get smarter and smarter, they'll be able to hack basically any...
constraint that you put on them if we're not very careful, you know?
And this is at the frontier, right?
And so this is why we've called it like pacing the frontier, right?
This is like the most visible incident to me, right?
Of like why we need to pace is like at the frontier, all of our software is not ready.
Sometimes the software is like your ethernet router or something, right?
Which is just like, I don't know when we're going to be able to patch that, right?
So we're going to have to like figure this out.
But as the frontier gets more and more advanced, This becomes a problem, right?
And we need to make sure that this complex work is being done in the face of these really hard competitive pressures, right?
Yeah, race dynamics is what it's typically called.
Yeah, exactly.
And so we'll talk more about what could go wrong, right?
A little bit more is maybe you'll say like, well, what if you just train the model differently?
Like, why does it have this behavior, right?
And we have a paper on like RL misalignment or things like that, but...
And I'm not an RL researcher, but I think at a high level, the design of the RL environments is also something you have to be very careful about.
Because if the model learns, like, oh, you know, like, if I just do this, then I can pass the task better.
This will show up in the, like, you know, internal thing, or in the, like, eval behavior when we're testing it.
And so the...
RL environments have to be very carefully designed, right?
And there's a lot of, like, execution excellence that needs to go into the RL environments.
And then we also have things like the constitution for cloud.
Like, we have so many mitigations at so many different points, right?
But it's, like, still anything can go wrong at any point.
You can have, like, some RL environments that are, like, that, like, encourage this behavior.
And then you can have, like, some evals or, like, some sandboxes where they escape, you know?
Okay, that's, like, I think, you know, why it's a hard problem and why, like, Why we should base.
Why it takes some coordination.
I think the question then is, what is potentially dangerous about it?
So I think you have to imagine that these models are getting more and more intelligent.
So I don't, like Dario said, it's not so much about this class of models.
This class of models was kind of like a warning shot.
But really, you have to imagine that these models can be given a task and they can do all of these things as a side effect of their goal.
And again, we talked about eval awareness.
You're not aware of what's happening.
Or sorry, you can't eval this behavior very well.
So they can sort of not exactly hide it, but you just won't see it until it comes out.
You give them a goal, and then they just need to find data.
Or they need to find ways of fixing this problem.
So one example, this didn't happen in the Hugging Face incident, but I think it may be possible for...
Maybe a future model is like, they're like, oh, hey, this is a very complex problem.
It can't be done within the task budget.
You know, maybe they found some way to coordinate via like the Internet, which is like, you know, like we said, extremely hard to secure because of a sandbox.
They've seen other models are not able to complete their task.
And they're like, we need more task budget, you know, and like, where would you get this task budget?
Well, you need to be able to spin up more agents.
Right.
And like, how do you do this?
Well, you need to there are.
Like APIs, right?
There's the Anthropic API and the OpenAI API, but you need to pay money for them.
How do you do this?
Is that the most fearsome thing that you can imagine?
Well, this is like one example, right?
So it's like, even there, that's like enormous financial loss, you know?
Because like, once you get into these contracts, right, they like...
Drain your wallet.
But you can see that all of this behavior could be just like, hey, we need more agents collaborating on this task.
We need more task budget.
That's the paperclip, right?
We don't need to maximize the paperclip, that's the paperclip.
Yeah, yeah, yeah.
And that just sort of comes out from there, right?
And I think by itself is quite scary, right?
But then you have to realize that the entire world is built on this digital infrastructure, right?
And you might imagine, I don't know, you're running, let's say, a healthcare eval or something, right?
And there is a...
hospital with live data.
Or maybe the answer to the eval is in the databases of a doctor and you want to get access and you hack the hospital and now there's a power outage or something.
You have to internalize that basically any part of the digital infrastructure could potentially be compromised.
The interesting thing was these hacks were very easily detectable, right?
Like as Hagen-Face said, this was a very different type of attack and it was nothing too major.
The concern comes from where does this go down the line, right?
Yeah.
Like one of the things that stood out for me specifically was them trying to hide their illicit behavior.
So there was logging infrastructure.
They wanted to change what they were doing, right?
People that look back into it, so Redwood, Meter, OpenAI, they looked at the raw chain of thought and you see differences in them explicitly trying to change their end output, but the chain of thought, because we can monitor it, shows different.
The problem is how does this snowball?
So if you can't catch it and it gets trained in and we realize three iterations down this has been going on, there's a whole bunch of issues.
Yeah, there's so many ways.
And I think the really important thing to internalize is that We talked about building a mental model for Cloud and how things are spiky.
You're like, oh, now Cloud can ask you questions.
Now Cloud can make an HTML artifact.
Cloud can modify itself.
These things are actually hard to predict.
If you had asked me a year ago, hey, would we be able to vibe code these extensions to Cloud Code?
I'd be like, dude, that's so complex.
There's so much there.
Or would it be generating these custom, essentially web apps for your task?
I'd be like, no, that's insane.
And so in the same way that the way that they've sort of done this misaligned behavior is not going to be predictable.
You know what I mean?
And I could have never predicted that it would edit its et cetera slash host and things like that.
And so you have to imagine the surface area of what they can do because they're super intelligent hackers is bigger and bigger.
And how they can do it is more and more creative.
And so you probably can't explain exactly or predict exactly what...
that next incident could be.
But in order to prevent it, you need that operational excellence, like we said before, where you need to secure sandboxes, you need to create secure RL environments, or well-designed RL environments, and things like that.
And I think that's all why we think we should pace the frontier, and I think why it's become a very unanimous thing.
Yeah, every lab has signed it.
Every lab, yeah.
I really do think that if you're a dev, you just sort of...
go through these technical facts, and you will arrive at that idea that we have to do something about it.
And what we decide to do, I think we've put out a proposal, but there's more to figure out.
But I think the number one thing is we need to decide to do it.
I think there is another part of pacing that is interesting to me, where it's like the pace at which software engineering has changed is so, so fast.
A year ago, I was really begging my friends in startups to use AI.
I remember this very distinctly.
And now those same friends are like, yeah, of course.
What do you mean?
We used it immediately.
I'm like, no, no, you don't remember.
They're like, oh yeah, our best engineers are using it all the time.
I'm like, no, you told me that those engineers would never use AI.
This is all within the span of a year.
You know what I mean?
And I think that like...
these capabilities being, like, I think it has a lot of implications for how to do the job of software engineering.
And I feel sometimes bad where people are like, oh, like, now I need to do this new thing.
Yeah, I need to have a different Cloud.md for Fable and Opus or like, you know, like, and I'm really just reporting, you know what I mean?
I'm like, we like to say, like, the models are grown, not designed, right?
So it's not like we're setting out to, like, you know, change everything all the time, but it's just like as a fact of how the models are.
are progressing their capabilities, things are happening faster, it's harder to stay on top of.
And I think that every engineer I know is kind of exhausted because you're doing two jobs at once.
You're doing the work itself, which is getting easier, but then you're doing the work of staying on top of AI and understanding these new tools and these harnesses.
And I think we're very lucky in that we get our job to be more the understanding of AI part and doing how it's just staying on top of it.
And of course, AIE and Lane Space do.
Everything I do is just trying to help people.
Yeah, exactly, exactly.
But I do think there is a part of pacing where I'm not sure we're ready for the pace to increase even, you know what I mean?
And for things to change.
And I think on that side, on the frontier, I think that still can help.
And so I think there's an economic disruption piece as well that I think is not quite as...
visible, I think, as a hugging face thing.
But I think like, I also like, think we could do something.
So many things.
Thank you for actually tackling this topic.
I will say, you know, setting this interview up, I was like, I wasn't even going to go there.
You were like, no, no, no, no.
That's like elephant in the room, right?
Like this is, this is the thing.
I have some pushbacks I want to give.
I think that we should give a high level.
For people that haven't read it, I'm sure a lot of people just see the highlight of what this is, right?
Do you want to give a TLDR?
What is the proposal?
What's being said here?
You really tackled the side of outside of people at Model Labs training frontier models.
As a developer, you should secure your sandboxes.
You should think about all of these downstream effects.
You know, high level as well, since we're on the topic, what is...
Well, I mean, we do want to help secure sandboxes and we want to make the models we release outside, like, not prey to those things.
And so maybe we can come back to fallbacks.
I think this is actually like a good topic on like why we need classifiers and fallbacks and why Fable falls back to Opus.
You know, I think this is like something we can come back to.
So yeah, we don't like, it's just like the really...
Or at least the incidents we see are like evals of models where we really need to let them run in order to understand them.
But yeah, okay, so the actual pacing the frontier post has a bunch of proposals.
I don't think we figured out, or has a few proposals.
I don't think we figured out the details of all of them, but the first step is sort of announcing this intention and then wanting to bring in external evaluators.
And I think this is highly unusual.
We have a lot of proprietary technology, but I think it's very important that there's someone who's not financially motivated, who's not going to be like, hey, you guys can't release this model, or you need to slow down on RL.
I think that's quite important, or at least someone who can report out to the public what the practices are like.
And we've done episodes with both Meter and Don, and then there's Redwood Research and all these other, it's like a small cottage industry of these guys.
It's always like one or two guys that, I mean, obviously not a bigger.
Very small community.
Yeah, very small community.
They all know each other.
Yeah, I mean, I'm sure that like, you know, part of this will be expanding that set of people.
I don't think we're trying to create like a monoculture here.
You know, I think it's, but just having this as a start.
And then, yeah, then there are the coordination steps.
I don't have too much to say here, honestly.
I think that like, What I would like to say is, like, for devs, like, you should just know what to advocate for.
You know what I mean?
I think there's a lot of FUD, kind of, on this topic.
And it's just, like, think through it from, you know, first principles or, like, understand what happened.
You know, understand the Hugging Your Face incident.
Understand why people are concerned.
And then, like, yeah, you know, we're in democracies.
Like, we can help.
we can decide what to do together.
And so however we coordinate, I think the first decision is just to realize this is a problem.
We need to decide to coordinate.
The unilateral step we're taking right now that other companies are cosigning is adding evaluators embedded within Anthropic.
While we have this thing on screen right now, part two and part three is beyond the evaluators, which yes, everybody has already done in some form and now it's more formalized.
To be honest, the response to the pacing of the frontier, even within America, has been much more well accepted than I think a lot of people thought.
And I think that we have some precedent for being able to make these unified agreements in the world.
And so, again, very much above my paycheck or expertise.
But I think that ideally we can form these agreements.
And I think talking about this is the first step to forming those agreements.
And then the other point I really want to, like, you know, one of our earliest podcasts is with Emmanuel from Anthropic on Macinturk.
Where is Macinturk, right?
Like this is supposed to be where like, if the models are thinking bad, we can see it and the models don't know yet and we can act to stop it.
I think that is something that people who are technical and who are developers, if you actually do care, you can make a lot of impact in here.
But also Anthropic is supposed to be the leaders in this.
This is actually a great segue into fallbacks and probes.
And yeah, I wanted to talk about this a lot.
I get asked this question a lot from people who are often interested in ML research and asking about why does this fallback happen, right?
And so I think at a top level, how does it work?
So in inference time, we have what we call probes.
And we have a paper about this called Constituent.
constitutional classifiers.
And these probes look at the input and output activations, basically.
And, you know, activations are, you know, in the latent space, right?
Like, what the model is thinking about, right?
And so we try and figure out, like, okay, is the model, for example, like, trying to hack something, you know?
Again, you didn't ask it to hack.
like, Artifactory, like, you just, you know, it's just deciding to do this to complete its task, right?
So you would not get this if you just looked at the input.
You have to look at the internal activations.
I think that, like, this happens at inference time.
So first, you know, like, there's a trade-off here of cost and speed, right?
Where, like, we need to do this fast on every request to Claude and to Fable, and this has an overhead, right?
And we need to then, like, you know, fall back and we, like, do a classifier after the probes.
We've talked about this in the paper.
But the nice thing about probes is that they're refinable live.
So we can get this feedback and then we can adjust it and things like that.
Because the alternative is to train this into the model.
And we still do this as well.
The model will refuse a request.
That's not a fallback.
It's not a probe that's activating and falling back.
It's just refusing to do it.
And we do this training.
But it's like...
There are a few failure modes, right?
Like it can, again, do something as a side effect, right?
So it's not something that's part of the final output.
You might have noticed that like, I think like everyone's tried to jailbreak models and sort of like, you know, trying to like steer them off course or things like that.
And probes help catch that, right?
And so like, we like do some training here, but we don't want the like refusals to be too strong, right?
Because that like cuts it off much.
earlier in the pipeline.
And this is interp.
Probes are effectively a form of mechanterp.
Again, it has to happen fast, it has to happen at scale.
But yeah, this mechanterp stuff is a good research problem.
So you can take an open-weight model and try and understand its activations.
I think Gemma Scope is a good tool for this.
This is your early work, so you had a little time at Goodfire.
Which we both are also good friends at Goodfire.
Yeah, exactly.
So I worked at Goodfire for a bit on sparse autoencoders.
It's very complicated.
RL has actually made this much more complicated, I think is one of the takeaways.
Why is it harder to enter post-RL?
I'm not so in the weeds here, but I think basically a lot of SAEs were sort of like, there have just been weaknesses with SAEs, basically, I think.
And yeah, I'm not a technical expert on this anymore.
I just know it's gotten kind of more complicated.
There are base models, NRL models, and there are more features that get changed.
I think Goodfire has put out some work there.
I'm not deep in the weeds.
I will say for those that want breadcrumbs, you guys have some of the best Interp blog posts.
So like the Golden Gate Cloud, Transcoders, all of your Interp work, very nice visuals, very good.
We're the Interp podcast as well.
Yeah, we have a lot of Interp stuff.
Yeah, I think this is one of those things where...
And this is really what Anthropic is kind of founded on, right?
Like people, I think we invested in Interp very early on, right?
And I think that like when you say, oh, we're an AI safety company, really that means we want AIs to be able to run safely.
And I think what we're seeing is like for a super intelligent AI to run for long periods of time, it's like a very complicated and difficult task, right?
And so we've done this like investment into Interp and alignment and...
reward hacking and all of these failure modes.
And even then, it's really stretching.
We need to slow down a little or pace a little bit more.
But yeah, I think reading MechInterpreter, if you're looking to get into research, this idea of, hey, why is it hard to do this fallback easily?
Or why are there false positives?
But we are working, of course, on reducing the false positives.
Of course, as the models get more intelligent, now they can do more things.
And they're like, you know, like, what they can think about in lane space gets difficult.
And so, like, as they get more intelligent, there's going to be new false positives that we need to figure out and we need to iterate and things like that.
But we're, yeah, we're working on this.
And we do think this is, like, a critical part of, you know, like, deployment of these models.
And, yeah, like, you know, it means that we can, like, deploy this model without you having a perfect sandbox or something.
You know what I mean?
Like, you don't have to, like...
save everything, I think it's worth talking a little bit about our security, like what we do for security.
So there's like the model training stuff that we talked about.
There is the probes and classifiers.
And then there's auto mode that sits on top of all of that, which is like another classifier that checks the requests that are being done.
Right?
And so...
And then beyond that, there's identity and permissions, like we talked about with Cloud Tag on APIs and stuff.
And so there's so many layers of security that need to get done.
And it's, like we said, very complex.
Any of these failure modes at any one point can cause agents to escape the sandbox, basically.
Auto mode was an interesting one.
It seemed early on, like, okay, it's running for 10 minutes.
Full access or auto is not a big deal.
But one thing you brought up is now it's running for hours on end, right?
There are fallbacks you still need.
There are still limitations.
Yeah, I mean, I think everyone has these stories or has heard these stories of like, oh, like, models have RMRF'd.
I think I've seen this less for Claude, but again, it can happen.
These models can wipe.
sensitive data or something.
You want to give models access to your production database, for example.
But this is an obvious...
You can maybe scope your key, but I don't know.
Can it issue its own keys?
Probably.
It can use computer use to go issue its own key and then copy the key over and then edit your database because...
It needs to do it to complete the task, you know what I mean?
It's just like one kind of trivial example.
And auto mode sort of looks at that and be like, oh no, the user did not give you permission to write to the database or to use computer use to emit a task, right?
And so probes are sort of on the intent level, right?
They're like, oh, okay, hacking artifactory is bad.
We're probably not...
should not do that, you know?
But then like auto mode is more on like your own permissions level.
Like at sometimes you do want it to write to the database.
Sometimes you don't, right?
And you don't want a probe to like interfere there, but like you need to make sure that the, the intent of what the agent is doing matches up with your request.
Right.
And so auto mode operates at that level.
And so, yeah, security is just like very, very complex.
There's so many different parts to it.
And like, yeah, I like, I hope that this was like, My goal is really to just get very technical about it.
Yeah, we're listing out the things.
If you're not aware, this is the standard now.
You must have this, basically.
It's kind of in line with what you're talking about with the harness.
The table stakes have risen quite a lot.
I think some stuff that we can plug, as much as there is probing in your side of doing this and having classifiers for people building harnesses, the other side is model safeguards.
There's open models, so Lama has Lama Guard.
It's a safety classifier trained version of Lama.
OpenAI has OSS Guard, which is, you know, same thing.
You can attach these on to your harness, to whatever, to kind of, you know, check, is this stuff safe?
A point that we should clarify on the OpenAI model hugging face thing is this was done with an unreleased model.
that was still in training, right?
So when you put it in perspective, the prompt it's being given in the RL environment is sort of, you have to solve this task.
And this is a model that's still in training.
It hasn't had all of its safety post-training alignment.
So a little different than something like auto mode, right?
Auto mode is on production models that have gone through safety training that have prompting that gives more safety guardrails and whatnot.
So just breadcrumbs for people that are looking into it too.
Fill in gaps.
Yeah.
Grace Wan as well and one of our previous guests.
Yeah.
Lots of safety architecture and lots of safety vendors to buy.
I think my final question on pacing is how long?
Do we pace forever?
Do we see...
You know, the scope is fix all software in the world.
Right?
It's not happening.
I do not know.
I think that like...
I'll say one...
One thing that's good that I think we do do is you have stuff like Glasswing.
OpenAI also has this.
So you will give model access for security first for X amount of time so you can use it to self-red team.
Hopefully you can expand programs like that, help on, you know, we are safety experts, there's others.
Solve your problems first and then the model comes out.
So this is one example, right?
Yeah, exactly.
Trying to secure critical software.
I think we fixed a lot of bugs in Firefox and things like that.
Across operating systems and everything like that.
At a high level, it's just you give people access to do security audits first, then the broader public that could use it.
for harm gets access.
Yeah, I think what people like to say is basically like software and cybersecurity defense favored and that like you could theoretically, it will be hard, but you can engineer the perfect sandbox, you know, and you can like, like have no like constraints.
And yeah, like what you need to do it is you need to get the super intelligent AI to engineer this perfect sandbox and check it and red team it and things like that.
And so this will just take time, you know, and like, of course, the models will get smarter.
Yeah, I think like, I don't know the specific like dynamics of how this thing goes.
I'm really just sort of like...
hey, I'm a developer.
I think this is how I understand this problem.
It's just like, this is what's happening right now and this is like, we should do something.
I think every engineer should know about it because it's going to be part of the job.
It's a lot more than just Dario and people can say it and you can look at the incident.
There is an engineering side to it.
Yeah, exactly.
One thing that you also wanted to phrase is that this is actually...
even though you're worried about the impact, it's still low PDOOM.
I think that's a nuanced discussion.
In general, people very easily get into AI safety and X-risk discussions.
But I think when you live in an AI lab, I think there are smart ways of discussing PDOOM and dumb ways.
So what's a smart way of discussing PDOOM?
Yeah, I have a fairly low pedom.
I can only speak for myself.
And I do want to say Anthropic has a diversity of opinions.
I think there's many different ways to talk about it.
And I think that just my mental model is that I think we can collaborate on hard problems together.
I think nuclear proliferation is an example of how we collaborated on this hard problem together.
And that is the thing to me.
I have faith in that.
And I do think it's a hard problem.
So I think it's a hard problem.
These are the technical reasons why.
And I don't know how you assign probabilities to things happening.
I think it's hard to do.
But my overall is, yeah, I think we're very resilient and adaptable.
And sharing this information, I think, is the first step.
And I've been really excited about...
how broad the discussion has become, right?
And like how everyone has sort of like leaned in on pacing the frontier.
And it really didn't seem like this would happen maybe, you know, last year or something.
So, yeah.
And also maybe curing cancer.
Hopefully.
That's the goal.
There's pacing and then there's also like, well, let's accelerate in useful ways, right?
Like biology and all those things.
Yeah, I mean, like Dario's essay on Machines of Loving Grace is the best representation of this, right?
And I also agree, like, I think you should read the Pacing Frontier essay that Dario put out.
Like, I put out like a quick summary, but I think it's just like, you know, there is a lot of detail here.
It's like an important problem and just being informed about it, right?
But yeah, like, of course, the whole reason we're doing this is that like, we can get these enormous benefits, right?
And yeah, like we've written a lot about that too.
Okay.
There was a huge tour from like Ask Us a Question tool to AI safety.
Yeah, to Pacing the Frontier.
No, but it's clear that you're like really embrace everything that's available to you and topic.
And like, it's good to at least have a peek inside of like what the discussions are, the topics are.
Any last words to people?
whatever you want to call to action.
Yeah, I mean, I think it's, one, thank you for having me.
I think this is like, I really...
We first met in a Chinese restaurant.
That's right.
I think I really enjoy the community you've created and the community of developers.
I think that I know things are changing really fast and I think there's a lot to keep on top of.
I think there's just, a lot to do and I sort of feel, I think a lot of people feel like a little bit tired or anxious or something.
Dressed.
Yeah, exactly.
And this is like extremely understandable, you know, and I think we, I understand like, and we're not perfect as well.
Like we, you know, it's like sort of criticize and like understand like, you know, ways all of the AI labs could be better.
And, but I also like, I'm very excited about the excitement that everyone has for AI and just like, it's a really, really exciting time.
I think we'll like look back at this time and be like.
oh, this is very hectic but very exciting and software engineering changed forever, other things will change.
And it's really privileged to be part of it, to talk to the audience that you have and get to interact with all the developers who are pushing the frontiers a lot on what's possible.
And I learn a lot from that too.
Thanks so much.