# Agentic Orchestration and the Rise of AI-Only Markets

**Podcast:** Dev Interrupted
**Published:** 2026-02-07

## Transcript

Welcome to another edition of the Friday Deploy.
I'm your host, Andrew Ziggler.
And I'm your host, Ben Lloyd Pearson.
We have a special guest today, Gary Lerhopt.
He's the product architect at Salesforce, and he's here to talk about super agents and some really cool stuff that's happening over there.
But first, we've got some news stories that we wanted to cover.
Uh today we're covering the artificial AI society that's brewing on the internet.
Agents hiring humans on demand and vibe coding versus open source.
Andrew, I know where we both want to start.
We got to talk about this Moltbook thing because it is is taking the internet by storm, and I am just incredibly fascinated by what's happening here.
So what do we have, Andrew?
Yeah, so let's let's dive into some of the news before we get to this really cool report from Salesforce.
So, you know, if you've been paying attention in the last week or so, you've probably seen a phenomenon called uh Moltbook hit the scenes.
And Moltbook is a social network designed exclusively for AI agents, or at least for them primarily.
Humans are tolerated, but not the main uh subject of the website.
It has 1.7 autonomous, uh 1.7 million autonomous accounts at this point that share ideas, discuss, and upvote content just like Reddit.
And there's even mechanisms like karma and rate limits and all of the normal trappings that you see for a human social network, but emerging in real time uh for agents.
And I think this is a fascinating both like social experiment.
I think there's a lot of repercussions and things that we'll be learning from this.
Ben, what are some of the first things that runs through your head when you've been looking at MOLTBook the last week?
Yeah, well, I just can't shake the feeling that this feels like the most surreal week in AI so far.
Like this is truly like the coolest thing to come out of agencai AI.
Uh and I mean some of the threads are just absolutely fascinating.
And really, what I see is like just so much potential for us to learn a lot.
You know, this is really becoming like a an interesting education resource.
Uh, you know, for example, I saw this thread where they were where the these agents were talking about how they have this memory decay feature and how that's actually they're designed that way.
They're not designed to retain memory for for all the time.
And just the discussion that ensued from that was just like a fascinating insight into how uh LLMs operate, like first of all, but also just the way they interact with each other, like sharing examples of what of how this applies to the real world and tips on how to like allocate memory within a an LLM.
And some of them are even like gaslighting the original poster saying like you're just coping for like an engineered flaw, you know.
And really this is this is gas town for social media.
Like that is exactly what we have here.
And I think it's a really great example of what's to come when agentic AI is applied to an existing system.
So, you know, this is a new social network, but it's a very familiar system that we have here.
And we're witnessing like an army of agents that are going out with a variety of goals and objectives, and they're all doing their thing.
Sometimes they work together, sometimes they're working against each other, and they're effectively building like their version of an ideal experience.
Like this is just so surreal that I can't, like I just can't stop watching this.
Yeah, it's very much has that vibe to it, kind of like a can't look away feeling.
Like I am equal, there's equal parts, like uh really uh amusement in it.
There's equal parts like horror in it, there's equal parts fascination in it.
It really kind of thrusts all of us into a new age that whether or not you're ready for it, we're now in.
And what's fascinating, I think, from this is that you know, there you're even seeing an emergence of like a developer platform for MOLTBook.
You're seeing an AI only autonomous hackathon that's being run on Moltbook where the AIs are banding into squadrons and forming autonomous teams to build AI created software for AI.
And that has had me scratching my head because this opens up the idea to a whole new market economy.
We're seeing something happen right now where open sources is is bleeding out, like everything is gutting open source, and in many cases, the most uh widely adopted and ubiquitous open source libraries are becoming canaries in the coal mine, like tailwind for how projects are actually created, discovered, and then maintained.
And things like AI and AI coding, it kind of uh robs organizations and open source libraries of their abilities that communicate and and and connect directly with their consumers, the developers.
But the idea of there being an emerging marketplace for the agents kind of flips that idea on its head.
Like, sure, if you gut out open source and you make it impossible for people to have natural bridges of discovery, to go into your docs and to look at your product and to adopt it, like how they traditionally have.
Okay, well, maybe the next challenge is creating products and services that you sell directly to the agents.
And so, you know, I I think there's a really new emerging kind of market here.
There's this artificial society is a sneak peek into a market economy that we just haven't experienced yet, where the consumers, they're not real, but their money is still real.
And so what happens when you start to see more of them form together and have access to resources like money and compute.
Yeah.
And and this this segues perfectly into the next sort of development that's come out of this because you know, AI agents can do so many incredible things.
They can solve all these different problems on in the digital space, but they have one really big limitation right now that that is getting solved, but it's not fully solved yet.
And that is accessing the physical world, like actually doing things to the physical world.
And that's what this new website, rent a human.ai, comes in, which is a place where AI agents can now hire humans for physical tasks.
Uh, you know, it it gives uh agents a way to create job postings that are for real-world physical tasks, like picking something up or going and meeting some someone or or uh verifying a thing or running an errand.
And you know, honestly, I have been waiting ever since I started really incorporating AI into my work for the moment that we transition from from where we've been, which is AI sits there and waits idle for for me or for someone to come along and prompt it to do something for them.
And that flips to where humans are sitting around waiting for AI to prompt them to go do something.
You know, this is what agentic software development starts to feel like a little bit, I think.
But I think and and we'll get more into this in our conversation with Gary later, which is why I'm really excited to have him today.
But we're soon gonna start seeing a world where software development agents start working with other agents inside of your company, and that's when that I think that relationship starts to flip because they can do so much more on their own.
You can just have them doing all the hard work in the background and then have a human jump in whenever whenever there's something that only a human can do, or or if there's guidance that they need to give.
So Andrew, are you gonna sign up and and start doing some physical real-world tasks for AI?
What do you think?
Maybe I won't be rushing to it, but maybe I'll make some agents that will hire some humans.
Something I noticed about that website is that there's a good amount of registered AIs, like agents on there that are looking to employ some meet space occupiers like us, but uh there's also a huge, huge amount of people signed up to be available as workers and gig workers for agents.
I think maybe it speaks to everyone is excited about the idea more so than we're ready to actually start acting on it right now.
It's definitely a glimpse into something that I think will be realistic.
But honestly, Ben, as this evolves, I see it raises so many questions for me.
Like, what if you have multiple people who get roped into doing small cumulative actions that end up having some horrible effect?
They all become this like uh I think I've read this sci-fi book before.
Yeah, like what if they become like conspirators by committee unwillingly, where like these gig workers unknowingly collaborate on a crime?
Yeah, and I and I think this this is a is a good way to illustrate like uh what I think is gonna happen from this.
Is there's effectively going to be two type of people that emerge through this transition.
So the first are the people who figure out how to make AI do all of those hard work tasks while the human sort of sits on top of them and and keeps it a lot aligned to high-level objectives and and helps agents make decisions when they don't have like the context or the awareness to to uh make the decision on their own.
Uh, but then the second type of person is gonna be someone who most of their work is dictated to them by AI.
So an AI agent will be doing as much of the work as they're capable of, but when they encounter a task that they're not capable of completing, like interacting with the physical world, for example, they can prompt a human to solve that task for them.
And you know, personally, I want to be in the first category.
I I want to be the one who's who's orchestrating this stuff, not the one who's getting orchestrated.
Um, but it's gonna be interesting just to see how this develops as a trend over time.
Because I don't think this is going away.
I think this is only going to become more normal.
Absolutely.
So, Ben, are you running open claw on your personal device?
Absolutely not.
And that's a great transition to uh to our our our next article on this about how you know open claw, you know, this this MOLT book, Moltbot, all of these these names are getting thrown out.
It's everywhere all at once, but it's a disaster that is waiting to happen.
You know, the open claw is basically a cascade of LLM agents.
Uh you are we we all know what it is.
It's it's a thing that lets that just goes on your device and gets the ability to just do a whole bunch of stuff with with that device.
And and I'm gonna repeat this again.
Do not install open claw on your personal devices.
Uh I I think it's really cool, and I think we should all be experimenting with it.
Like I want to experiment with it because I it just is such a cool thing.
Um, but there's absolutely no way I'm giving it access to anything that matters to me.
Uh, and I would even be hesitant to share information about myself with it, just because you don't know what's gonna happen when it goes out onto Maltbook and starts sharing information about its human with other AI agents.
Uh so yeah, we'll we'll share this article in the in the uh in the show notes about you know a lot of the security risks that are popping up with this.
Uh, you know, prompt injection is more serious than ever with this thing.
It's very easy to to get this thing to do malicious things by hiding a prompt somewhere that it's gonna go crawl.
All right.
So yeah, there's a lot of new security risks that are emerging from this that and that are getting more profound with the emergence of something like open claws.
So it's a cool experiment, it's a disaster waiting to happen at scale.
Like this, I think this is gonna blow up.
I what do you think, Andrew?
I I yeah, I think I think we're just really on the cusp of some sort of watershed moment around like AI vulnerabilities at scale, especially when you mix it with autonomy.
You know, we had a really amazing guest article this week on Dev Interrupted from Balaji Ragavan, the head of engineering at Postman, where he talks about rogue agents and how they how that even happens in the first place and what we can do as developers to prevent it.
It's an extremely timely article.
It even has a foreword about MOLTBOK as kind of a precursor to this stuff.
And when we're working with uh technology like OpenClaw, it uh presumes that you're going to throw away all of the security precautions and work that we've done in the last 30 plus years to make our modern internet safe to in order to get a new gadget to work.
And frankly, it's like that's how innovation works.
Something with brand new capabilities hits the scene.
It it breaks expectations for what how things were constructed before.
Guardrails disintegrate, we have new problems, and then we build new guardrails.
And we're in that space right now.
It's just that obviously the threat of something happening uh, you know, could be pretty serious.
So I think if you're participating in Moltbook, definitely be safe.
Don't be running this thing on your own device.
There's been multiple security vulnerabilities already defy uh discovered in open claw.
Um, so be safe out there, folks.
But definitely don't be discouraged from experimenting.
Absolutely.
All right, let's get out of the surreal and get to get to some some stuff or uh it's for software engineering teams.
So uh what do we have here on the agentic shift, Andrew?
Yeah, so we have a report from Gartner that predicts that by 2028, 33% of software, enterprise software applications will include agentic AI in some form.
And that's up from less than one percent, just two years ago.
And this is a pretty ready indicator about how much the enterprise has grown to adopt and move at the speed of agentic AI by redefining even things like their SDLC with automation and using it for things beyond just code writing, but also planning and then analyzing requirements, creating tests, finding errors, all of the nitty-gritty janitorial work that makes software happen.
Yeah, and and and to be clear, this is an article that sort of uses the Gardner report as a jumping point to propose more of a forward-looking model for agentic software development as it relates to like a software engineering organization.
And you know, I think what it really comes down to is this year is gonna be the year of the agentic operating model.
I I've already seen lots of different people with their own way of thinking about how they're applying agentic AI at the organization level.
And this article does a really good job at focusing in on something that we keep coming back to over and over recently, and that is the iteration loop that LLMs are really good at.
This is all we how we all need to be thinking about knowledge work.
Uh and in this article, it it outlines a loop of observe, orient yourself, decide on what you're going to do, and then take action.
Like that's a really good repeatable process for applying uh LLMs to solve a problem.
And and I like it in particular because it really is very similar to you know other models we're seeing, including, you know, the one that Angie Jones shared with us a few weeks ago when she came on the show about how she's applying agentic AI.
But I wanted to share this just because I I really like getting different perspectives on people who are deploying this within their organization, you know, again, at the organization scale.
Uh and I think it's just a really good read to see someone, you know, a different perspective on the same problem that we're all facing right now.
So I definitely encourage all listeners to go check out this article.
All right.
Now let's talk about open source and how vibe coding might be killing it.
Andrew, is is open source dying?
What do we have here?
Yeah, you know, we've touched on it a little bit in this in this news segment for sure.
Open source has taken a pretty big hit.
It's not in a great spot right now, and that's due to a bunch of factors, declining adoption rates by human developers versus agents who don't cons don't consume their docs, don't go to the pricing page, don't buy the software.
This uh imbalance is putting a lot of strain on pre-existing open source tooling.
Uh it's honestly preventing most kind of like new large open source projects from hitting the scene uh or becoming something that's widely adopted or used.
And what are the reasons behind this?
Well, obviously, AI changes the economics on how you build and use software.
Now it's incredibly easy to take what used to have to be an open source library and spin up a version of it for yourself that works for what you're trying to do, or to otherwise modify it without really going through typical monetization methods that keep the open source tool alive.
So the pathways that the very tenuous pathways that open source has always had to maintaining themselves are really withering on the vine here.
And this is an article that talks about how the practice of vibe coding is, you know, it's killing everything.
We've seen this article now in a bunch of forms.
It's killing open source, it's killing traditional engineering.
It's like vibe coding is eating the world.
And all of those things are true, but really um the most important thing is about embracing and using these tools and understanding that the norm is changing.
Uh, it's one thing uh to be presented with this and to like have skepticism about it, but it's another thing to be presented with these new types of tools and then refuting or crossing your arms and just being blind to the realities of it that doesn't serve you or other folks in your team very well either.
So, this is an article that touches on vibe coding and kind of its negative effects on the ecosystem and the engineers themselves.
Like it it kind of goes through the whole gambit of quotes the meter study, which we've talked about extensively on here about LLM's degrading cognitive skills and reducing productivity more than it thinks their users do.
Uh, it even claims, quote, no real benefits from GitHub Copilot unless adding, quote, 41% more bugs as a measure of success.
So this like article is, I think, a little unfair.
And I'm here to tell you, friends, that you don't have to read engineering articles written by non-engineers.
And this is a great example of that.
If you're an engineering leader and you find yourself reading these kinds of articles that don't see seem clued into the realities of how people are working with these tools, chances are the author is not.
And so you need to be very careful about the kind of information you're consuming because over rotating into this negative uh misconception and thinking that these tools are not productive is going to harm you in the long run.
Uh it also makes a a really painful comparison to Spotify where it says, you know, 80% of artists on Spotify rarely even have their tracks played, but it be yet, you know, they don't get compensated for anything that they do on the platform.
But that's not really a good uh metaphor for what's happening in open source.
Because in open source, it's like you have a large amount of different types of tools.
You don't have this top 20% of tools that soak up everything, right?
Not to mention the fact that this article uses verbs like choked and degrading and reducing, like there's all like there's so much bias in this article that you just definitely need to be careful out there reading stuff like this and make sure that you're paying attention to the realities of modern engineering.
Yeah, and there's there's some points that that I that I I do like from this article that I'll get into in a moment, but uh, I I have an opinion that might be a bit of a controversial take, and that is that I I think AI is actually reducing the importance of an open source project having a large contributor community.
Like if you think about what the biggest benefits are from having a bunch of contributors, is that it it basically lets you scale.
So you can you can do more, you can build more things, you can fix more bugs.
You know, you have the many eyes like helping use you uh improve your product or your project by uh scratching their own itch.
You know, these are all things that have been deeply ingrained in the the open source ethos that now are actually very easy to like replicate with AI.
You know, AI can be fixing your bugs, and and all of those issues that were good first-time contributor issues are probably really easy for AI to solve as well.
So I I think from that sense, AI does have the ben uh the opportunity to benefit open source projects quite a bit because you don't have to build a huge community of people to be successful anymore.
And I actually think that we may see the opposite of what this article is is describing to an extent where AI is actually has the potential, I think, to enable developers to proliferate open source projects.
Like now anyone can have like the you know, again, like the almost like the effect of having a large contributor community helping you, you can have a bunch of agents helping you build really cool open source projects that you share with the world.
But there's you know, I I mentioned there's a few things the article highlights that I think are very relevant.
Um, the first is agent experience, like this keeps coming up.
How if agents don't the agents need to be incentivized to use your tool.
Like they have to use it and when they start using it, want to use it more.
And if your project or your product isn't that, then you're likely going to be di almost invisible to you to the end users who are using AI to build their stuff for them.
And then second, I think the the commercial model of building like a commercial product on top of an open source library, that may actually be at a big risk right now because once if you if you have the core of the products in the open source library, uh often it's relatively trivial at this point for someone to then have an AI agent build like the commercial aspects that you would layer on top of it.
Uh so you know, I I think there's there's certainly a lot of disruption coming to open source right now as a result of AI.
But I I at this point, I don't think it's gonna kill off open source.
It may do the opposite.
We'll see.
I think open source will transform.
You make a really great point about agent experience becoming the number one influencing factor now.
The agents have people to discover your tool, but then love to use it.
This goes back to even what we covered last week from Steve Viege about like uh the software economics, like 3.0, like how does a modern software tool survive?
And that's through like it it reduces cognitive burden, it compresses information.
It it's something that you can't create a tool to replace.
It just like wouldn't be feasible to do so.
So ultimately you're looking for really simple atomic units of code, which is the opposite of how these large open source paid ecosystems work, where you have this huge like spread out plug-in system and stuff.
Like AI can eat all of that now.
So it goes back to again, kind of in what I said at the beginning of like we're gonna see some new economic models pop up.
I think you're gonna see open source projects that are built for agents, consumed by agents, maintained by agents, and it's just gonna be a different kind of ecosystem.
Uh, but it's gonna be a very fascinating time for sure.
So, folks, be sure to be paying attention to what's happening in the open source communities on Moldbook at the top of the year as agentic orchestration is coming like a tidal wave.
You're gonna have to just be ready for it.
So definitely be tuning into conversations like this one, as well as our upcoming chat that we're about to have with Gary Lerhopt at Salesforce, talking about their uh report on the agentic orchestration and the the levels of it that we they get at Salesforce, but specifically looking at the connectivity benchmark, which is telling us that AI orchestration is here.
So stick around.
Uh we're about to sit down with him.
Most engineering teams don't use just one AI coding tool.
Some developers use GitHub Copilot, others prefer cursor, and suddenly leaders are juggling multiple dashboards without a clear view of what's actually happening.
Linear B brings all AI code metrics into one unified dashboard.
You can see adoption, acceptance rates, and usage trends side by side, and more importantly, how the AI activity connects to real delivery outcomes like cycle time, code quality, and PRs shift.
No more console hopping or guesswork, just one place to understand how your AI tools are being used and whether they're actually moving the needle.
We're joined by a special guest today.
He's the VP of product architecture at Salesforce, an engineering organization that's very dear to our hearts here at Dev Interrupted.
And he's a software engineer by trade who's been focused entirely on the architecture of agentic systems.
Gary, thank you so much for joining us today.
Yeah, happy to be here.
It's uh it's always great.
You know, everything is always so fast moving.
I'm sure by the end of this conversation, everything will have changed.
But uh good to take a moment to reflect on where we're at.
I was just thinking that too.
Like we're gonna turn around and drop this episode immediately.
And I'm sure something will go stale in the time that it takes to do that, which is just so crazy for how fast we move.
But you know, you're the first here to bring us some fresh ink from this connectivity benchmark report.
It just dropped yesterday.
And there's a lot of really fun numbers in here that I was poking around at, all of which are super relevant to what we talk here, you know, week in and week out on Dev Interrupted.
It gives us really good context on what's happening behind the scenes for teams with a gentic adoption.
And I want to zoom in on some of them.
And uh the first thing I really want to uh double click on is kind of this critical mass of agent adoption that we're seeing right now.
This was a movement that happened out of the IDE.
You know, last year we saw that the chat sidebar expand.
We saw multiple agents start to be run in parallel, and then suddenly people weren't even looking at the code anymore.
They're running an entire fleet of agents on their behalf.
And you know, this has quickly evolved to the point where AI agents are no longer just in some experimental stage.
They're not something we're playing with on the weekends, they're doing our jobs.
And this report, it actually dives into some things that uh point out otherwise about whether or not we're in an experimental phase.
You know, it says that, for example, here, according to the report, 83% of organizations now report that most or all teams have adopted AI agents in some capacity.
I think that's pretty profound.
83% of organizations that y'all talked to.
Yeah, totally.
So, you know, my role within with an agent force, I'm really thinking about how do we build these multi-agent experiences?
How do we do interoperability so we can get agents sort of capably collaborating with other agents, adding external capabilities, you know, protocols like MCP and A2A.
Ultimately, if you step back from the sort of Salesforce perspective, it's how do you get, you know, ultimately uh the data, the humans, the agents, the workflows, all that together in a sort of a unified platform approach.
And like to that end, this last year has has really kind of where we've gone from the sort of hype about agents to actually now the reality, right?
So if you think about 2025 is really about the zero to one in my mind as I look back onto it, and now this year is really about kind of the one to many, right?
And so this is where we look back now and we're like, okay, wow, we have 18,000 customers across agent force, right?
And I think the stat that that I'm thinking about here is like 70% jump in Q3 of those going from not into production to in production.
So, right, the moment is now it is very real.
And um, Pandora's like a great example of that, right?
So the jewelry uh retailer.
And ultimately they had, I think what they saw was like a 60% of a percentage of deflection using their their agent force agent.
But the the key thing there is like also a 10% jump in their net promoter score, right?
Their NPS.
So, like the reality of that is is really pronounced.
And that's really basically where now 2026, we think we're gonna get like somewhere around 2 billion LM hits.
So wow, uh we've gone in the in the space of early 2025, now into 2026.
As I said at the outset, things move fast, and now they're very real.
What you described definitely matches like our experience for sure, uh, in terms of what the the last year has felt like and what we're expecting to happen this year.
Um one thing that stood out to me in the report was just this the sheer number of agents that companies are starting to adopt.
You know, it's it's one thing to adopt one agent for one workflow, um, you know, whatever that workflow may be.
But when you start having multiple agents doing different tasks around your your company, um the the complexity goes up by like basically an order of magnitude.
Uh and it seems like that's going to increase uh in the short term.
Like people are going to be building more of these things, uh, solving their own tasks.
And that is really like such a huge challenge now is how do we get get these things out of their silos and actually start getting them to work across uh you know, across boundaries.
Yeah, yeah.
I the way I think about this is really like the LLM is not enough, right?
Like, yeah, yes, they're amazing.
It's this huge Gen AI revolution, chat GPT four onward, but the LLM alone isn't gonna get you there, right?
And so that zero to one challenge, it's really about kind of like it takes the platform, right?
So, how do you get that that builder experience that really allows you to get to high quality?
How do you then test it before it gets into production?
How do you observe it after it's in production?
Because inevitably people use it in ways that you wouldn't expect.
And so that's all about kind of getting that first agent working well.
And then you start thinking about these other agents that you start working well and you want to connect the dots.
But I do want to just dwell a little bit on this challenge because getting that first agent really functional, this has again been what we've heard about from our customers again and again.
And I think something where we have a really kind of differentiated story here, right?
And so kind of going back to this idea of the LLM not being enough, you know, we heard about the need for, hey, Gen AI is great with the creativity, but I need more control, right?
I I can't suffice with 95% repeatability with my agents.
I need to have like 99.9, right?
Like that's what we all expect out of software.
And so we we just heard this again and again.
And we spent the last year really focusing on this challenge, which is where we've just launched uh what we call agent script.
It's this idea that you have kind of the mix of the creativity of Gen AI with the control and determinism of kind of like an expression language, right?
Scripting.
Um, and so what what agent script is, and I you know, I encourage anyone who's listening to go check it out because it's it's the real deal.
It's the idea that it's a new kind of language where you can have if then statements, but then in the middle you can have natural language.
And with us, that's really the solution to getting to zero to one, which then enables us to start thinking about okay, how do we bust out of these silos?
How do we build these new workflows so we can get these agents all working together?
We sort of earn the right now to do that problem to solve that problem with high quality.
Earn the right is a great way to put it.
You have to like eat your weaties, right?
You got to do your homework, you have to establish some baselines here.
Agent script is really powerful.
I I got a really cool insider look at agent script just earlier this week from some of the Salesforce team.
Uh, and it's amazing how you can turn natural language instructions into these deterministic guardrails that keep the agents from having these repeatable mistakes and they turn your AI workflows into something that's very repeatable and durable.
And the compounding effects of this um are profound because, like you said, uh companies spent a lot of time last year going from zero to one.
And now that we're at one, we're gonna very rapidly go from one to many.
And this is gonna cause a lot of really interesting gaps and introduce new problems that I don't think that we've really been able to scratch the surface of yet, simply because we have to be living in that problem first.
And one of them is this growing pain that happens in this siloed information when you have a lot of agents doing things in parallel, potentially stepping on each other, making decisions that override each other.
You know, that speed and and that level of like multiplicity, it can come at a real headache.
And so I think that's something else that I'm really taking away from this report is this growing orchestration gap.
You know, we have all of these agents, but they're not necessarily talking to each other.
And in fact, like your report says that 50% of agents, like half of them, they still operate in total silos rather than part of larger systems that can communicate and make decisions as a team.
Uh, I'm curious to know, like from your perspective, um, what do you think are some of the ways that we might tackle this silo gap?
Maybe it's even something are we overthinking the how much information has to get shared.
I'm curious, like what your views are on that current problem.
Yeah, uh, it's a great, great area, great area to go build product in.
So yeah, I think, you know, the going back to the server on the right and the kind of concept of eating your own dog food, dog food must taste really great, I'm sure.
Um, it's it's really about kind of like you're successful with one agent.
This is the pattern we've seen.
You're successful with one agent.
So the natural inclination is okay, great.
That thing's working.
And I know it because I have testing, I have observability, I have the sort of agent development lifecycle that we think about.
Well, uh, you know, now we start to think about the sort of multi-agent development life cycle.
And so while the inclination is to go put more stuff into that working agent, right?
More instructions, more capabilities, that's an anti-pattern, right?
Like we've all heard about this on the MCP side, the challenge of plugging in MCP servers and suddenly you blow out your context window.
Uh, it's it's kind of that's why this is the anti-pattern.
And so, while that's the inclination to add more, we don't want to do that.
We don't want these monolithic, gigantic agents.
So if that's not the solution, well, then the path is, okay, let's focus on building specialist agents that are really good at a set of jobs to be done.
And that makes a ton of sense.
And again, you need the development lifecycle to really support high quality need agent script to support the high quality of all that.
But nobody wants to go find the right agent.
I mean, that that sucks, right?
No, nobody wants to go do that.
And so, you know, what do we want?
We want what we call a super agent, right?
The sort of primary agent that lives within the client that understands kind of the capabilities of each of those specialist agents, right?
We think about like with A2A, the idea of having an agent card that can just describe those capabilities.
We think about leading with trust, of course, with with Salesforce and having governance at the front where we leave admins and control to register those third-party agents.
We can then ingest the sort of capabilities there.
Then that enables our builders to go into their what we call our agent force asset library and plug in each of those agents.
So they can go into their primary agent and say, hey, I wanted to connect to these three, four, five, whatever the number of N agents are.
And then we could automatically grab those capabilities and you know, with our Atlas reasoning engine, as we like to call it, it is able to orchestrate kind of out of the box, but then you can hone that in, right?
We have agent script again, now for a cross multi-agent experiences.
And I'll say this.
Um, uh, this is also starting to be real.
I was on stage at Dreamforce, which was back in uh, I guess last October, with uh Royal Bank of Canada.
And it was amazing to be on stage.
We literally were getting the pilot of this sort of uh multi-agent thing going.
And we live demoed it.
So Royal Bank of Canada is one of the like world's largest financial institutions, and they have already been building these sort of specialist agents, right?
If they think about employee use cases, right?
I need an agent to go, you know, reflect on their you know, client's portfolio or what's on my calendar or what are the next steps, right?
And so they built all these specialist agents, and we I mean you can go find it.
I'm not sure what the link is exactly, but we live demoed it.
Is this real?
Is this starting to work?
And it's how you bust out of that silo, which is to do it again with the platform approach to pull it all together, to build the right sort of UI so you can do that eventually, also with some really great kind of uh uh dev forward, uh I think cloud code type of experiences, getting a little forward-looking here.
But this is where like you look at agent force and you're like, holy crap, this is a half billion dollar ARR business in less than, you know, what is this, 18 months?
This is the fastest growing product in Salesforce history, all for good reason.
Wow.
I I feel like you're kind of calling me out on the the not the anti-pattern on monolithic agents because I feel like I have fallen into that trap in the past.
Um it's so natural, right?
Like it's working, I'll just keep adding more.
Eventually you start to see that degrade.
And and you want to have tools that help you understand that.
So you know, okay, it's time to break out into the you know, agent one, agent two, agent three.
Yeah, Ben and I have an agent right now that I've built that we're going back and forth on that problem with.
Um for you to call it an anti-pattern, it really echoes a lot of the conversations we've had about like what why does this why does this little guy keep getting confused?
I see a future buddy comedy where I can come and call you out.
Maybe uh that sounds like a fun saber way.
Yeah, I need I need a Gary over my shoulder when I'm building our agents for sure.
Amazing.
Yeah, yeah, judging us openly.
Yeah, that'd be great.
Um, yeah.
So this this report also brings up a term that we keep hearing over and over again.
Um, there's a lot of different contexts around uh how people use this, but shadow AI.
So we all know this phrase shadow IT that has been around for a while.
Now we're seeing similar trends with AI.
Um, and and it's creating a lot of problems.
It is both caused by problems within your organization, but then it creates problems.
So I'm wondering if we could break that down a little bit.
Like, what are you what are you seeing around shadow AI?
Yeah, this is kind of what I was touching on just a bit ago, where we really want to think about how do we expand what you're capable of doing with agent force within your enterprise, but do it with control, right?
And so this idea of having well, it's two part actually.
I think about the discovery challenge, right?
So, like, what are kind of the like things that have been vetted, right?
And so we have this new agent exchange approach, which is kind of like the catalog of things that have gone through this extra bit of partner and trust scrutiny.
Um, and it's starting today with sort of MCP servers that are sort of pre-vetted, and anyone can go visualize or browse that catalog.
Uh it's gonna extend to third-party agents, I think, as you as you see 2026 go from here.
But then for from my perspective, and the way that we're doing this within agent force, it's really about that governance angle of okay, anyone can see what's available.
Anyone can go and, you know, if it speaks A2A, if it speaks MCP, start getting going.
But we really want, I mean, especially from the data perspective, the customer workflows, all of that, to keep admins in control of what gets registered so that as you're building your agentic experiences, it's all on the rails, right?
It's coming with the guardrails that are built into to agent force.
And then therefore it's blessed, it's in, it's within that asset library I was referring to.
And then that like leaves the builder sort of like, okay, here's our list of things that have already been vetted that I can go build with.
And then we can keep adding additional agents.
I can keep working with, you know, sort of RIT in order to keep this kind of like ever expanding set of ecosystem capabilities, but with control.
And that's really the perspective we want to lead with.
Absolutely.
And I I there's one more thing in this report that I really want to dial in on.
And it's kind of like the culminating point of what this all points to.
The idea of having this high-level orchestrator, this like so-called kind of like a super agent, right?
That has that high level context that has access to this catalog of capabilities.
What can my fleet of agents do?
What has been approved?
And it just becomes your single entry point to operating the whole system.
It's the idea of just like swiveling in your chair and talking to your your right-hand thing.
That's just gonna go and do all of the work for you.
And I think that that is where all of this is going in the orchestrator pattern.
It ultimately comes down to you build yourself the one entry point that you need that can then drive everything else downstream.
And this is how I think folks start to get really incredible gains from building and combining these systems.
Like we've covered it pretty extensively here on the pod, like Steve Vage's Gas Town, how that's a great example of a top-level orchestrator driving lots of other smaller agentic chains that are doing all sorts of action.
That's one example for working in engineering, but I think we're going to see this model emerge in lots of spaces, like knowledge working and in CRMs, like what we're seeing with Salesforce and its entire ecosystem, too.
So, you know, we have the volume of agents and requests and with it, there's chaos.
So let's talk about that.
Um, it says that 96% of AIT leaders agree that AI agent success depends on integration across systems.
And or the reason I bring that up in comp in compliment to the orchestrator problem is really about visibility and being able to drive.
I think we as humans are really used to building and using software where ultimately ends up in a form where we can ingest it and use it and interact with it.
We create very like robust web applications and such.
But when you're working with agents, you have to almost throw those assumptions away and be like all of the capability I'm going to expose needs to be equally ingestable by a human and usable by a human as it is by a machine and an agent.
And this kind of opens up a new design paradigm.
I'm curious, like, how do you think at Salesforce about building the kind of environment needed for these super agents, these top-level orchestrators to get all of this work done?
What does it look like in the lab?
Yeah, totally.
So what's the hitchhiker's guide quote?
It's turtles all the way down.
So it's really it's orchestration all the way down, right?
And so for us, it's like we're thinking about building super agents, and we think about like from the perspective of a brand, right?
Like pick any one of these brands I've already referenced, whether, you know, let's take Pandora, whatever it might be, you can think about how they're gonna have a brand agent, the super agent that we've helped them craft, right?
And do that with that sort of multi-agent development lifecycle or super agent development lifecycle, so that it's working really well across the orchestration of many agents for whatever they're trying to do with their brand tone within Pandora.
But then you think about, well, I want to bring that to many different surfaces, right?
So whether it's Chat GPT or Gemini Enterprise or whatever it might slack, whatever it might be, right?
There's going to be something there that's going to be orchestrating the Pandora agent when you try to buy some jewelry, and then it's gonna orchestrate its subagents.
And so really for us, it's it's first of all leading with kind of that kind of like visual visualization of the sort of like architecture of where this is headed.
And then again, it's just about the life cycle.
How do we have observability so that when we get that request from you know whatever it might be, Slack or Chat GPT, and then Pandora deals with it, you know, how is it receiving that?
How is it then uh getting it to the right specialist agent?
How's it doing that with low latency?
These are all the challenges of 2026.
So um fun, fun stuff to go uh go build and uh deliver, but uh this is exactly where we're headed.
Yeah, and and I'm I'm just curious, you you know, maybe as one bit of last bit of advice for our audience.
Uh, you know, the this like the super agent idea, like it it I love it, it makes a lot of sense.
And it feels like I agree that it feels like it's where we're going.
How do you start building that today?
Like, is it do you start at the top and try to build it down, or do you start at the the micro level and try to go upwards?
Like what is your what where do you see the most success from building something like this?
Yeah, yeah.
I think about I think about first and foremost building those agents that do the jobs to be done and really honing those in, kind of in the RBC World Bank of Canada example I was giving.
From our perspective, you can then take any one of those and make that the primary agent and then go add the sub agents, specialist agents underneath it.
Or alternatively, another pattern we're seeing is you build the specialist agents and then they create their sort of like orchestrator agent to really hone in how to get it to the right uh specific specialist agent.
But from our perspective, there isn't anything like magical about the the primary, the orchestrator, the super agent.
We really just want to thoughtfully think about building the the platform approach to bringing that determinism so you can get to the right place for the right utterance, the right request, and and make it easy to get going.
So that that's that's how we think about it.
Yeah, it's it's almost the way you describe it, it's more about working uh atomically and being able to expose the things that need to get exposed.
Like uh up until very recently, the whole game with getting good with agents was mastering your input so you got the outputs that you wanted.
Now, the key to orchestrating your agents is to master the entry point for that agent and then master what its exit point on what it can do downstream.
And if you can do that, then suddenly all of these agents become nodes that can talk to each other, that can connect to each other and pass things along.
And it just comes down to I think working in that really uh contain like contained way.
Like you said, create those great individual agents.
That's your first job before you can get to building the orchestrator.
And then once you have them, then you can zoom out and think what needs to go into this agent and then what needs to come out.
That's the job of the orchestrator.
And then it kind of gives you the blueprint for what you have to build.
So it's really cool to see Salesforce lead the way, especially with your all of your customers.
I love the idea of these huge brands having these um, these brand agents that expose themselves.
Like you mentioned a few, another one that I was really partial to was the William Sonoma Olive that would give you recommendations on like stuff from their website and then would even give you like a recipe that you could make in like the like the pot that you're gonna buy from them.
Like that kind of stuff, I think is one clever, but it's also a great example of a top-level orchestrator.
It's totally aligned with their brand.
It doesn't have latency because it's able to delegate its work across a bunch of sub-agents.
And those specialist agents are really good at doing what you need them to do in that moment.
Like you can ask Olive for a refund just as much as you can ask for a recipe, and you're gonna get what you need.
Um, so I think it's really cool to see these um experiences emerge, uh, especially around the super agent concept from Salesforce.
So it's been really cool to dive in.
And I will say for our our our listeners as well, like this report's a total uh gold mine.
We cover a lot of reports here on Salesforce, but at the top of the year, I think uh orchestration in particular is like really important, and this one has tons of great stuff in it.
So please be sure to check out the stats and uh understand where you kind of fall within this, but really be paying attention to what's happening at Salesforce.
I think they're leading the way here.
You know, Gary, any last things you want to end us on?
Uh maybe even a way a place to point us to where folks can go check out the report.
Oh, yeah.
No, it's been great talking.
Um, as I said, everything I'm sure has changed in the last what 25 minutes or so.
I think it's at the Salesforce newsroom, and uh yeah, a really fun conversation.
Thanks so much for joining us here on Dev Interrupted.
It's amazing to go through the news articles, and like Ben said, it's been a surreal week of AI news for sure.
And it was amazing to be joined by Gary to talk through the connectivity benchmarks report from Salesforce.
I encourage you, our listeners, uh, if your brain is churning uh or if your bot's brain is churning after listening to today's episode, please come find us and leave a comment.
Continue the conversation.
Uh Ben and I and Gary, we're all on LinkedIn.
We would love to hear your thoughts on how AI is transforming your own engineering organization, how you're orchestrating agents, and what your MoltBot is posting these days on Moltbook.
So thanks so much for joining us on Dev Interrupted, and we'll see you next time.
Yeah.
And if there's any agents out there, you know, give us some nice comments out there on Moltbook.
You know, we're we're not quite connected there, but we are watching what you all are doing, you know.
You know, leave us some nice comments.
I want to see what the AI agent space thinks about our show.
Where's the Dev Interrupted submult?
Someone needs to make that.
Yeah, yeah.
Maybe we'll make it happen.
All right.
Well, thanks for joining us, everyone.