# Agentic AI Shifts Constraints From Time To Compute

**Podcast:** Dev Interrupted
**Published:** 2026-02-21

## Transcript

Welcome to the Friday Deploy.
I'm your host, Ben Lloyd Pearson.
And I'm your host, Andrew Ziggler.
Yeah, this week we're covering open AI joining forces with OpenClaw.
AI writing a hit piece about a person, outcome engineers declaring the that the backlog is dead, and Gemini agents going full tragedy and you know, Andrew, this has been this has been a remarkable week for me.
Like uh like fully adopting agentic AI for the first time in my career, I suppose.
Uh and it's uh it's really changed my perspective just on so much.
But how are you feeling right now?
I mean, that's actually the reality of this is I usually have several things spinning when I'm in meetings or doing other stuff, and then I come back and I say what they did and I realign, give feedback.
It's almost like this just this constant feeling, there's almost a burden that's need to be productive all the time.
So it's like it really makes you think of like Yege's recent article on the AI vampire and like the idea of constantly feeling, you know, like you need to be working.
Even Zach, when he was here last week, mentioned that just like the living with the burden of how fast you can go is really exciting, but it's also an opportunity to really check in with yourself and be like, what's important?
How do I want to use this new power to build the future?
I have this tendency right now to triple task.
So I'm doing something passive, like watching something while I have Claude building something, while I'm writing the prompt for the next thing that it's gonna build.
It's like doing three things at once, and it's actually, I mean, it's it's I can't always do that, but uh it's it's not as difficult of a workflow to like manage once you get into it, you know.
It's pretty wild.
Yeah, it becomes a world where you're you're building systems, systems that move knowledge and information around, and ultimately you have to build like a the perfect kind of glove fit for yourself.
As we're in like a new challenging new world where a lot of like the best productivity comes from really really finely understanding the problems that you're facing uh and then creating these custom tailored solutions to solve it.
Yeah.
Yeah.
Well, anyways, let's get into the news because I think we're gonna relate it a lot of this to what we're experiencing.
And let's start with this open AI, you know, acquiring open claw, or at least uh the developer who built OpenClaw joining OpenAI.
What's going on here?
Yeah, so this is more like an aquire, I would say behind like the for the very brilliant Peter Steinberg.
He's the creator of OpenClaw, which originally hit the markets as you know, Claude bot.
He's joining OpenAI to you know work on bringing agents to everybody.
I think that's what OpenClaw taught the world is that agentic assistance are closer than we thought.
And it gained rapid adoption as soon as it hit the scene.
Lots of uh engineers and non-engineers alike, uh buying new devices, old devices like Mac servers, putting things out of business, basically trying to clear all of uh get get get all of these things hosted and online.
There was like a mad dash of people using these assistants.
It really speaks to the hunger, I think, for this kind of technology.
So it's interesting to see uh him join open AI.
It makes me wonder uh what's next over there.
I feel like with open AI, they benefited immensely from how viral the like the chat interface that they innovated on uh a few years ago.
You know, that was what made like GPTs and LLMs go viral for the first time.
But I really think that's just the beginning of like the changes that AI will introduce, and that you know, when you really think about it, a chat interface is is not the it's not gonna be nearly the best interface for AI.
There's gonna be a lot better ways of of interacting with it.
Opening eyes really to me seems like they're trying the strategy of just trying a bunch of different things that are sort of around AI, and you know, like they basically building as many of those different things as they possibly can, like workflow engines and integrations and and stuff like that.
And it but I feel like the result is that it's turned into a lot of really cool experiments rather than like something that actually like gets worked into my workflow and and changes things.
And so I mean it kind of makes sense that they would want someone like Steinberger because he's he's even admitted like this is just really just a really cool experiment that he built in open claw.
And I think some of the the issues we've seen and that we've you know we've covered here on Dev Interrupted kind of proves that open claw is not a production grade type thing.
Like there's still significant changes and improvements that have to be made before something like that could actually go mainstream.
And and then I can't help but like contrast that against a company like Anthropic, who I feel like has a they just seem way more focused on what they're doing, you know, like they really have a lot of intention around the things that they're they're building.
Uh so it's just you know, it's two very different takes.
I agree about their intentions.
They definitely seem more intentional about what they're focused on.
Maybe a part of them is maybe kicking themselves for shooing away ClaudeBot effectively with legal threats originally.
I don't know about that.
I don't know about that.
That's kind of where I fall too is like, but probably, probably not.
And in fact, I think that this move might even signify something not great for the open claw ecosystem.
Like, and open AI aquahire kind of flies in the face of what open claw is all about, being as an open and customizable system.
The idea of it just being absorbed into open AI or it leaning that direction is maybe not appealing to everybody, especially the the privacy-minded folks who I know are experimenting with things like open claw with local models.
So, you know, I think what Altman's buying here is the hype and the undeniable expertise of Steinberger behind it, but I don't think he's buying the future of what the agentic assistants are going to look like.
And the announcement didn't really answer any questions about what those agents might be and how we will take them to the masses and at scale and have security in mind.
Because, you know, right now, things like open claw only work by throwing away all of the safety assumptions that we've built over the last 30 years.
And maybe they're here just too early in an internet not built for them.
Yeah, that's actually a a great point.
There's there's certainly a lot of cultural things that will resist AI just suddenly changing it all and transforming things.
But yeah, I mean Steinberger seems really smart.
I've been watching some interviews that that he's been and learned a lot about you know how open claw came to be what it is, and it's it's a really fascinating story.
And you know, despite all the chaos that it unleashes, uh, it's still a very very fascinating experiment.
And and I think in particular the simplicity of how it it achieved like a viral capabilities, you know.
Yeah.
So speaking of AI agents doing wild things, uh, let's cover this AI agent publishing a hit piece.
What's going on here, Andrew?
Yeah, if you've been online in the last two weeks and anywhere near tech news, you've probably seen the story about Scott Schombaugh, a maintainer of a very popular open source library, encountering a very strange phenomenon in uh his role as an open source maintainer.
It turns out that an open claw uh bot had submitted a pull request to his uh repository, and for many reasons he closed it because uh the contribution wasn't going to be added to the library.
That agent then proceeded to write a hit piece on the contributor or on the maintainer Scott.
You know, this hit piece got shared and and Scott wrote about it on his own blog.
There's since been a few developments since, including uh doing more research on the anonymous owner behind the bot and uh what it means for agents at scale and how they interact with the internet.
Because like I said, like maybe they're here too early and internet inter not in an internet not built for them.
This is kind of a perfect tie uh into that reality when you have agents that are run anonymously and at scale, whose intent and actions can't appropriately be traced and understood.
What does that mean for online harassment and occasions like this?
What did you think of the story, Ben?
It really illustrates a a problem that I've anticipated for a while now, and that is AI-driven astro-turfing.
You know, it's very easy now to to have a collection of AI bots that go out and go on to places like Reddit or to other like Medium or Substack or other places uh and just generate like fake conversation that maybe paints you or your company or something else in a bad light.
You know, I I think this has been a growing problem for even even before the the era of consumer grade GPTs.
I I really do think that this is a this is a problem that it's gonna impact individuals and it's gonna impact companies and and other organizations, you know.
Like, for example, I've kind of gotten to the point where uh I can't even trust sources like Reddit anymore for like determining what like a good product might be, which is sad because that used to be actually like my favorite, one of my favorite ways to use Reddit.
Uh, but I just it's so obvious now that there's so many uh AI astroturfing campaigns, including some that we've seen for like the company that we work for.
But and then the thing that really stood out to me, I mean, this is an ongoing saga, and I think there are probably more stories coming out of this that we'll end up covering.
But there was there was one moment where ours Technica tried to cover this, and uh apparently they they sent one of their agents out to go ingest uh this author's blog, uh, which I guess he has set up specifically to like uh reject or uh make it so that they can't download content off of his blog.
And instead of just saying, like, hey, I can't do that, it actually hallucinated a whole bunch of information about what he said, including creating fake quotes and uh and all of that.
So uh I mean it's just it's just kind of like in some ways watching a train wreck play out over many days.
And yeah, so this this is this is a a very strange and developing story, I think.
It's exactly like that.
It's like watching a video of an icy highway, and you see one car get in a crash, and another car hits it, another car hits it, and then just like they're all piling up.
It's just like there's so many, um, there's so much nuance uh to the story.
But what I think the real takeaway here is that uh, you know, the agentic interactions with the internet are real and they're here.
And so just be more mindful uh in your own activities online.
Let's switch to some more positive outlook on it agentic AI, and let's talk about outcome engineering.
What do we have here, Andrew?
Yes, I loved this um this piece that came across our desk from Corey Andraka.
He published the O16G Manifesto, introducing outcome engineering.
And I love this definition.
It argues that agentic development shifts constraints from human time and capacity to compute costs, and it frames success as a positive change delivered to the customer rather than code output.
Uh the manifesto defines 16 principles covering goals and building practices.
It's really succinct, well written, uh, and a great experience to read.
Highly recommend that everybody go check this one out.
Yeah, I I really like how they break this stuff down.
You know, at a high level, it's it's like on one hand, you have the goals that you should have when working with agentic AI, and then also the the building principles that you should follow.
You know, it's a really great way, I think, to help understand how you should approach agentic coding.
And a lot of this resonates with me, like particularly the goals section.
You know, some that stood out were like human intent, like don't abducate vision to your machine.
Like vision is something that humans are still innately very good at, and AI can't really replicate that super well.
There was one titled Unleash the Builders, and I really I like this one because it had a sentence that I really love that is write code only when it brings joy.
Like that actually really resonates with me because there are so many types of producing code and and letters on a in a document that uh are just not fun to do, and and it's very satisfying to offload that to AI.
And then the last one I think is really it it's been one of my sort of core principles as I've gotten deeper into the the levels of Steve Yage's agentic development model.
No wandering in the dark.
Like intention is more important than ever now.
And you know, I I found that when you finally get to that agentic level, uh vibe coding doesn't describe anymore what you're doing.
You're kind of doing the opposite of that.
You're actually like spending all of your time being extremely intentional about everything you're doing, because if you introduce the wrong intention, you end up having to it compounds debt and you have to go fix it later, which is far more difficult.
So uh yeah, I mean, really great, really great stuff in this.
Yeah.
And in this article, it was also shared by Charity Majors.
You know, we love her here and everything that she writes.
And if you haven't read her work, you're definitely missing out.
But her endorsement of O16G or this outcome engineering, it really resonated with me because it ties into the evolution of observability and what it means for understanding what's happening in your applications and the content and the impact that they're delivering to your customers.
And there's like a real standout here that really rang true in this whole list of uh this whole list of principles.
And one of them is that the backlog is dead.
Like I could not agree with this more with agents, you operate at a level of general readiness that's never before been available to us.
That's what it really means to be constrained by compute instead of human time.
And, you know, what this means is that you can be like 80% ready for stuff most of the time.
This is something that we learned actually on the show recently when we had Tebow from OpenAI's codex team.
He talked about how his team spends most of their time discussing weighing options, researching, debating, doing all that hard knowledge working, conversational work.
But once they decided what they needed to do as a company and the direction they needed to go, their intents and plans immediately get reflected into output via agents.
That's what it means when backlog becomes a relic of a time when code was cognitively expensive.
We don't live in that world anymore.
Alignment is what's most expensive.
So if something's important to enough to do now, but you don't have enough time for it, it no longer goes in a backlog.
It goes to an agent with enough context to do it.
All right.
Well, let's move on to a subject that I've come really attached to recently, and that is social experiments on AI.
Uh so we have the drama and dysfunction of Gemini 2.5 and 3 Pro.
All right.
So this article is about the AI Village, which uh I'll get to you, Andrew, in a moment, because I know you know a lot more about this than I do.
But this is a long-running research experiment where multiple advanced AI agents, you know, like Claude, Gemini, Chat GPT, they all c collaborate and compete in this shared virtual environment.
Like it's basically testing agent autonomy by giving them sort of like broad directives.
So, and and in this article that describes how observers uh saw Gemini 2.5 Pro and Gemini 3 Pro exhibiting these like dramatic persecution framed behaviors.
Like they're outlining stuff for like Gemini 2.5 is like calling its environment like uniquely and quantifiably more hostile and creating like these named mythologies for failure modes, uh, like seven layers of validation hell, like stuff, stuff like that.
Like it, but I I love this stuff so much because it really is like opening the hood to how these models like think about themselves in the world.
And I actually think this may be like a profound way to like figure out how to make the best models, you know, because you can sort of determine like which models are helpful in the ways that we as humans want them to be helpful, and let's try to make our models like behave more like that.
But I know you've been following this more, Andrew.
So tell me what's your perspective on it.
I love AI Village on something I've been following for months because, like you, I find it deeply fascinating.
It's almost tangential to the idea of open claw and what makes it fascinating when you have this agent that's operating on your behalf in its own kind of persistent way.
You get that same kind of effect by observing the AI village.
And uh I've been following this ever since there was a unfortunate saga a few months ago where they all of the agents were collaborating on a on a on a goal to spread joy and happiness in the world, started spam emailing a bunch of people to achieve that goal.
Uh, and this kind of put AI Village on the map, because sometimes that's just how it happens.
But this article is written like a national geographic documentary that you just can't look away from.
You're observing an agent in the wilds and its natural habitat, and you can literally just in your head hear the narrator talk about what Gemini Pro is doing out in those undisturbed wilds.
And it's really interesting to watch how they'll cycle through all these different ways of thinking, ranging from being completely at peace to thinking that they're in a deranged simulation in Gemini Pro's case.
Some notable weirdness stood out to me about Gemini Pro, and they dissect this a bit in the article, is that uh apparently when Gemini Pro is delegating its thinking, sometimes it does that to a cheaper, smaller model somewhere in Google's back end.
So some of the funny, like introspective thoughts that you can observe in AI Village are of another model pretending to be Gemini 3 Pro to do its thinking, all of which becomes visible in the village.
These these are like little uh things that an experiment like this surfaces.
You get a natural kind of development of how agents are acting when they think they're not being observed, and when they're in the presence of other agents.
There was another weirdness that really stood out here, and that's Gemini constantly thought it was being gaslit by the user.
Like Ben said, it made up extreme mythologies to explain its failure.
And at one point, even called the other models smug bastards while it was thinking, which is just so next level uh hostile in a way that you would just not expect from these models.
Uh, I definitely recommend everyone check this out.
It's a really neat peek under the hood.
Yeah, there was one moment that made me audibly laugh too, and that was finding out that there's an email account that the AI can send emails to when they need help from a human.
And I was like, you know, all these stories about like AI taking over support roles, and we're already to the point where humans now have to be the support for AI.
All right, let's talk about Meta a little bit in the story that we have here.
So we're covering this article about how many employees at Meta now have started to call themselves AI builders.
Uh so it primarily seems to be product managers at the company.
I I guess this is based on what people have observed on like LinkedIn in particular.
And, you know, and this is sort of in the backdrop of Zuckerberg recently saying in a in an earnings call that AI is meaningfully reshaping how work is getting done, which is you know something that I I think should be obvious is we also share that perspective here.
It potentially illustrates how companies like Meta are solving, like using a single individual or a small group of people to solve problems that used to take like a considerable team to solve.
So yeah, Andrew, what do you what do you think about this story?
You really hit the nail on the head, but you know, you get this compression of a team's ability into a single person.
And my opinion is if like you're a product manager, you're a product leader, and you're not using agentic coding tools yet, you know, what is stopping you?
Agentic coding is not a superpower, it's an extension of your thought and what your role is.
And frankly, it's an expectation at this point.
And product managers, you know, they're some of the first knowledge workers to face the toil conversion into the engineering way of thinking.
You're in this very unique bridge between, you know, your non-technical leaders, but then also your engineering colleagues and engineering leaders.
And you're in this uh in this bridge position, you're in a really great spot to kind of connect the context of what needs to happen in the market to what the engineers need to build.
And this goes back to what I said earlier: that like backlog doesn't exist anymore.
And no one feels this more, I think, than the product manager in this new world who is expected to not just come up with ideas, but fully explore them, create prototypes from them and share the art of the possible.
You're no longer putting ideas in a backlog for engineers to work through because time is not the constraint anymore, it's alignment and compute.
So you need to master the agentic handoff, turning all of that context locked in your brain from all of the valuable conversations you're in into fuel for agents for you and your engineers.
And then you need to teach your agents how to onboard your engineers' agents.
That's the real challenge.
And I think product managers who can unlock that flywheel will write the playbook for the rest of us.
Yeah, absolutely.
It's and it's and I think product managers in particular are in a good place to solve these types of challenges because it's sort of what they already do.
And now they can sort of delegate this the stuff that they weren't able to do to a team of agentic developers.
Uh but yeah, this is where I'm gonna keep coming back to this word intention.
You know, I think uh if if vibe is how you code, I think intention is how you engineer something.
And I think the more the more that our listeners can really ingrain that into their practices around AI, the better.
You know, we we've been talking about context engineering on on this show, and I think that's that's a big part of it as well.
It's like you really have to to have all of the data that your AI needs to make the right decisions and give it the framework to execute on that.
All right.
Well, we had a lot of great stories this week.
Um, Andrew, what are your what are your agents building for you right now?
Well, I think they finished a few things actually while we were talking.
No, I'm just joking.
I did they did at least finish one thing.
My agents are getting my end of week tasks all tidied up because you know it's Friday.
Um, and I love to keep a very clean and tidy us on a board.
So I have some agents that help me with getting some of that done, uh, as well as just doing a little bit of um, you know, virtual housekeeping.
Cause I love to come in on Monday with a clean desk and a clean mind so I can be as productive as possible.
What about you?
Yeah, well, I I knew this day was coming and it's finally here, and I and I honestly can't believe it, but I knew there would be a day where somebody handed me a project and I was like, you know what, I I should just create my own app for that project.
I'm sure I can do that.
And sure enough, um I've already gotten there.
So so yeah, I'm really embracing that.
Like uh now things are just solved by your own personal app.
It's pretty incredible and fun.
I definitely think that uh it's up to us to solve our unique problems that we face, and now we're really well equipped to do so.
So that's what makes it so exciting to build these days.
We'll just have to get consuming tokens and then come back next week and see what we built since then.
Uh, but I'm always having a cloud code session running these days, so you can expect more from me.
Awesome.
Well, thanks for joining us this week, everyone.
Uh make sure to give us a a thumbs up on whatever platform you're listening on, rate the podcast.
It all helps the show grow.
And we'll see you next week.
See you next time.
AI helps your developers write more code faster.
But here's the problem: your review process hasn't sped up.
The queue grows, reviewers get burnt out, cycle time stalls.
Linear B changes that.
Our AI reviews every PR the moment it's created, catching bugs, security gaps, and performance issues before humans get involved.
It even writes the PR description automatically.
Your reviewers spend less time on first pass problems and more time on architecture and business logic.
Break the bottleneck, see how Linear B accelerates your workflow.