From CRM to Coding Agents: GTM as a System of Action with Swan AI

Julia Nimchinski:
Hey!

Amos Bar Joseph:
The second, Julia, Yes, can you hear me?

Julia Nimchinski:
Yup, and we are live!

Amos Bar Joseph:
Okay, amazing! Hello, everyone!

Julia Nimchinski:
Awesome! Welcome to AI Practice Sessions! And today, to everyone watching, we will explore the transition from systems of record to systems of action. We’ve prepared some amazing frameworks, AI-native workflows, and models to showcase this new paradigm.
As always, follow the conversation in the HOC Slack, ask your questions, we’ll make sure to address those, and traditionally. We always open those sessions with the black swan, himself, from Swan AI, welcome to the show, Amos Barr Joseph, long time no see. How are you doing? Congrats on the funding.
And, to everyone watching, it was… did something super untraditional. So he fundraised to not hire people, but to invest in agents. Tell us more about that, Amos. Let’s get into it.

Amos Bar Joseph:
Yeah, so always excited to be here, Julia, and I feel like, you know, from time to time, the expectations just get higher and higher, so, you know, hope to deliver as well today. Yeah, we are trying to build a different type of company.
For those of you who don’t know, we’re building Swan as an autonomous business, a business that is kind of designed from the ground up to scale with intelligence, not with headcount. we didn’t raise, hire agents, not humans.
What we did is we actually, we raised the funds to focus on the talent that already is in the company, instead of hiring more people, and just… stack up intelligence on top of them. So we always like thinking in autonomous business as having two types of intelligence.
One, which is the most important one, is the human intelligence, and then, on top of it, we stack artificial intelligence to support, you know, the talent that is within the company.

Julia Nimchinski:
Love it, and useful context for everyone who’s watching. So, you embody this new model of running a company to invest in humans, you say, but how many humans are there? And, yeah, you’re just basically living in the new model, the three-pounder type of, you know, unicorn. So, tell us more.

Amos Bar Joseph:
Yeah, so, in 2025, we started Swan in January, we launched the product, basically just 3 founders. I was doing growth, you know, anywhere from, getting the story out there to meeting with, customers and closing deals.
Then we had another founder on the product, and then another founder on the, you know, the R&D side, but all of us are kind of technical and kind of, super hands-on. And we grew from, 0 to over 200 customers, just the 3 founders, with a lot of help with our friends, the agents. And it was a crazy ride.
This year, we’re trying to do 10X, so from 200, customers to 2,000 customers, in the realm of the 7 figures ARR. I’m gonna drop some numbers soon for the folks who are following.
And we’re not gonna stay 3 people, we’re gonna get to 6 people, so we’re doubling the team size, from 3 to 6, proving that, you know, small teams, you know, can beat the giants, and we’re fighting the biggest companies in the industry. We’re getting, actually, heads-on into Claude, so it’s gonna be an interesting fight.

Julia Nimchinski:
That’s awesome. Well, let’s get into the topic of today’s practice sessions, and I… I’m really, really curious your thoughts on this one.
We kind of touched a little bit On, the strength that is happening on the previous practice sessions, and… I’m speaking to this, investment thesis by the majority of leading investment funds, like, you know, Foundation Capital, Vessemer, A16Z, the next trillion dollar opportunity, your favorite article out there.
Systems of action, context graph, agents, all that. And, it sounds like, finally, the premise of, you know, the identified GTM is happening, because you can obviously capture the decision traces, the exception logics. Every data point that essentially was… always lived in people’s heads.
So I’m curious, Amos, how do you see this architecture evolving? And in terms of… there are so many various playbooks and theories at this point. Like, do we just use a data warehouse, something like Snowflake and Databricks, and then just connect to Swan AI? What happens to the CRM? Do we get rid of it? How do you see it?

Amos Bar Joseph:
I love it. Yeah, there’s, so many different angles here right now, so many different voices.
I think the sad thing is that, you know, the majority of these voices are not… in the weeds, they’re not really the ones that are building, they’re not the ones that are, confronting these challenges head-on, and that’s why I think they look at it from what I call the ivory tower, you know, from, like, 10,000 miles above.
They’re looking at it, and I’m talking from, you know, ground… Leveling, no ground level, from the folks that are actually building with this technology, and I have a relatively different point of view on how, it should actually look like.
So, a couple of things here for the folks that, are just… either starting their way into this context, thing in the AI world and trying to figure out what’s, right and what’s wrong. So the first thing is that, data is not equal context, okay?
Basically, when you look at something like Snowflake or a CRM, there’s so much data in it, and there’s bad context and good context that lives there, right? So… you know, some things, are not updated correctly, or not updated at all.
You don’t want the AI to think, yeah, these companies, what they do is they don’t update deals whenever they get to negotiation. That’s what I need to do, right? It needs to understand That this is actually, you know, not following a specific process of the company, right?
And so, there’s a lot of these discrepancies, et cetera, and actually, context doesn’t live in the data. Context, most of it, and the majority of it, actually lives in our heads, okay? And, so that’s the first point. Second point is that to get context. into your AI, so that it’s able to work with it.
And for those of you who are not sure about why we need to get context into, you know, the hands of the AIs, because without context, without understanding how your business works, then AI is worth nothing, right?
It’s always… there’s this analogy of having the smartest person in the world, if it doesn’t know anything about your business, won’t be able to help you, right? And so.
The second point… is that to get the context into the hands of the AI, it’s not about machine learning, it’s not about, you know, technology that is super complex to deploy it, and do I need to drop, like, a large language model on my Snowflake instance to start having a powerful AI? No, it’s actually… That’s the work, okay?
It’s called context engineering, that’s what humans should do. You need to describe to the AI how a process looks like. And you need to put it in a file where the AI could always approach it, and make sure that whenever the AI needs to follow that process, when the timing comes, then it could access that file, okay?
We can talk about what’s the right architecture for that, aka coding agents and a file system, but the process itself That’s the thing that humans do in this revolution, basically. It’s context engineering.
We’re starting… we’re moving from doing all… what we call system engineering, like clicking on buttons and connecting workflows and trying to configure stuff.
to actually context engineering, we are the ones who are in charge of describing to the AI what’s good looks like, what bad looks like, what’s the process, where the data lives, where are these systems.
We’re giving the AI the context, and we’re in charge of what I call context parity, to always make sure that as the business evolves, as our GTM evolves, we always keep that parity intact, because the moment that we have it, we have something that’s called context drift.
Which we’re going to talk about maybe today as well, Julius, which is super fascinating.

Julia Nimchinski:
It is, and let’s dive into the coding agents, because there are so many definitions. We live in the ultimate age of marketing of AI.

Amos Bar Joseph:
Yeah.

Julia Nimchinski:
The latest one I’ve heard from A16Z is that there are two kinds of agents, essentially one that just does action for you, any action, and a second one that collects information for you. I know that you look at it a little bit differently, can you just share?

Amos Bar Joseph:
Yeah, I think that, we were kind of, like, in this interim phase where humanities tried to understand how should we build agents, and then, you know, in that interim phase, we had all these definitions, but I, you know, predict.
I don’t like to predict stuff, I always love to talk about, you know, today, and what’s working today, but from what I see that is working for us, and for a lot of, like, AI-native startups out there, I think, and I predict, that in the next two years, we will only have one definition, real definition, for an agent, that everybody will follow.
And that an agent is a coding agent. Okay? All agents must be coding agents. What does it mean to be a coding agent? Basically, a coding agent is an agent that can execute, can not just write code, can execute the code itself in its own sandbox environment. And it has a file system that can… it can write to that file system and change it as it go.
And these two capabilities, why they are, you know, the defining moment of AI agents, why we came to this powerful architecture with these two components, is because they share two important things. The fact that the agent can actually execute code within its own sandbox enables it to become limitless, basically.
Okay, it can do anything it wants, basically, because there’s no limitations. It can write the code. If it doesn’t know what to do something, quote-unquote, it can write the code and execute it on the spot. And also, it becomes kind of, like, digitally native in this software environment.
Instead of trying to communicate with words with all these software tools that we built around it, it can actually speak their language, right? It can communicate just like Neo and the Matrix can communicate with the entire world around it. Same thing applies to an encoding agent that can write and execute code.
Second thing is a file system That’s the memory, that’s the state, that’s the context, that’s where all the how-to lives. And that… the file system enables a simple mechanism for humans to tell the AI what to do, and how to do it, and when to do it, and for the AI to constantly iterate on it in a very simple way.
And so, what happens when you pair these two capabilities together is you get, basically, the next platform, the next destination, the next… Place where humans go to perform their work.
And what people are not used to yet is that they think that coding agents are something for developers that, you know, hackers can use for other work, but maybe we’ll talk about it today. Julia, we’re actually one of the first ones to use this technology to apply it for non-developers, not for software-related work, but actually for GTM work.
And I think that’s the future of all the platforms, is to either be a coding agent or die.

Julia Nimchinski:
That’s awesome, that’s an outtake, for sure. Look, we live in this, cycle of words, so last time we met, it was the context type of cycle. All the X feeds, YouTube, LinkedIn, just one word every post. Now, it seems like governance. And no one has the ultimate answer, but I just have to ask you, what’s your take on that?

Amos Bar Joseph:
Governance is a real problem, and… let’s… so let’s break it down, like, what’s… what’s govern… governance in the context of agentic work? So, there’s several things. One. There’s this simple thing that you now have endless power at the fingertips of every person on the team. Again, human-centric.
Don’t think that because it’s agents, everything changes, and we need to rethink about everything.
Same thing, humans using software, we need governance around it, the only thing that changes is that, wow, there’s, like, these… stochastic, probabilistic creatures that we need to create some governance around, and it’s not code, so the paradigm changes from, like, I can limit code to there’s, like, these token-based beasts that I need to make sure that a user that uses these token-based beasts doesn’t erase a database, right?
So that’s, like, one part of the governance layer. The other part of the governance layer Is around context. And again, it’s similar to what we had before, so before it was data, and you said, you know what, yeah, you know, the SDR can’t see what the CEO is working on an M&A deal, or something like that, right? Same thing goes with context.
When you’re starting to really build this context layer, and that coding agent can access everything, then you need to ensure that, you know, which type of users can use the AI to actually get to it. And again, because these are more, like, reasoning-based animals. So it’s harder to create something that is really bulletproof. In that sense, right?
And then… The final piece… is agent-to-agent governance, okay?
So, I do believe that over time, we’ll have less and less and less agents, and we’ll kind of, like, fall into this just one coding agent with a file system, but still, even in that future, we’ll have two, three agents working in our company, and… they’ll need to, you know, to share context and share tools with each other, and share work with each other, and so there’s, like, how do you govern that process as well?
Because each agent has its own permissions and its own, system access, etc, so you need to have agent-to-agent governance as well.

Julia Nimchinski:
Beautiful. Amos, if a mid-market or enterprise just, you know, I mean, obviously everyone has the mandate to rethink their tech stacks, GTM motions, everything. When I start to… this process of, identifying AI, nativeing, whatever you call it, what would you advise?
How do you even approach it, the transition from system of record to system of action? Do you get rid of it? Yeah, like, what’s the playbook?

Amos Bar Joseph:
Definitely. Yeah, so… Couple of things here. One, just… First tip that is not… super AI-savvy or something like that, but it’s very simple to apply that I think everybody would just, if they do it. So there’s, like, a simple predictor for, like, if you’re gonna fail in the AI implementation or not.
And it’s, did you start with the solution or with the problem? Okay, that’s, like, the first thing. Starting with, we need to move from, like, a system of record to a system of action, that is starting from the solution, okay? And if you start your thought process there, then most likely you’ll fail with your AI implementation.
What you should do is, like, okay, where do we have problems? today, in how we run our GTM.
And then… most of the lines will start being drawn to, okay, then we need a system that don’t only store information, it could actually be more proactive about that, and actually, you know, help collect that information and help act upon that information, but, like, what type of information? Which… in which areas, like, which processes, etc, right?
So if you start with your problems. then all of this will demystify much faster, like, how you should get there. So that’s the first tip that I recommend everyone is, like, again, nothing has changed in the software world. You still need to start with your problems, and then you get to the solution, okay?
But what we’re seeing in terms of, like, the tech stack. that should be built, with, you know, we see it at, like, mid-market companies that are more, like, AI-savvy, and, you know, bigger enterprise will follow later. That’s, like, usually how technology waves, proliferate. So… They usually have 3 components. One The CRM, still there, okay?
It’s a database, and it’s just less important in the stack, but it’s still the place where you keep structured data on your customer relationships, okay? And you spend a lot of money and a lot of efforts to configure it so that it has a good representation of how you would like to store that information. It’s not there.
But it’s like, it says, yeah, we have, you know, the close date, and maybe you have, like, a custom object about your, whatever, right? You have a lot of these properties that you kept, and you worked a lot on making it, look like it could be a good place to store data, just the data is not there yet, right?
So you still need that, and it’s good that you have that CRM. You’re not going to throw it away. What you need on top of it is two things. One, it’s basically a system of context and orchestration. Okay? So, that system can proactively listen and be present in every touchpoint. and make sure that the CRM is updated.
And it can work with other tooling in your stack that you have in your GTM, so if you have maybe, like, sales automation, or if you have, like, forms, or whatever, that system should be able to work with everything. So it does two things.
One, it is able to work with context that is not structured data, so it can actually hold We talked about that file system and that processes, etc. So it could actually hold, you know, unstructured data that explains how your organization works.
in a way that humans could interact with it and tell it, yeah, no, we do it differently, actually, yeah, that’s how we do inbound assignment, that’s how we route leads, and things like that. So it stores the context, and it has an orchestration layer so that it is connected to everything.
So it can apply that context And follow the processes to make sure that CRM is updated, that other systems are updated, that reps are notified, that we’re sending emails, that we’re doing all this stuff based on the processes that we encoded there, okay? That’s where SWAN, by the way, lives in that architecture.
So you have your CRM, which is, like, structured data. Then you have a layer that handles context and orchestration, which we believe should be a coding agent in a file system. And finally, you have a co-pilot, like Claude, okay, which is a different coding agent that you go to to do most of your, like, ongoing work. And so what happens… is async.
In the background, you have the context and orchestration layer running all the time. At all the different points of touch points that you want to have it. work, making sure that, you know, everything is following the processes that you would like.
If you want to do work right now, you go to that co-pilot of yours, and you want to do deep work, that co-pilot is also connected to a lot of your different systems, so you can access different systems over there. It should be able to talk to that context layer as well. And so, just by having these three components.
You can basically move from the traditional world, where you had a system of record that didn’t really have good record of how your business worked.
To more autonomous operations, where you invest more in the methodologies and in, like, how should we do stuff, rather than actually executing stuff, because you have a lot of autonomous Reinforcement that works for you.

Julia Nimchinski:
Love it. Question. when you think about all of the, I don’t know, 20 years, 30 years of, you know, GTM as it was, till this point. We had thousands of sales tax solutions, thousands of market tax solutions, CS, etc, etc, in every category.
And it seems like we are transitioning to the same kind of situation when we’re just basically identifying everything, and thousands of solutions are emerging. And just for syncing it all, it was from first principles.
when you think about system of action, and yes, you want to capture everything that is happening in your GTM, sales, marketing, CS, and just make it actionable and, I mean, potentially autonomous, that would be amazing. But when you think about that, what… how do you see this step, ultimately?
Is it a system of action for marketing, a system of action for sales and CS separately, or is it even on your roadmap to capture the full funnel, or… yeah, just if you can speak to this.

Amos Bar Joseph:
So, I’ll tell you how we work, and I think that that’s how I think the future would look like. So… just like Salesforce built, like, the platform initially to capture customer success, marketing, sales, and support.
All of these four verticals, they run on a shared layer, eventually in Salesforce, and a lot of organizations are running on, like, you know, all these instances on Salesforce itself. The way that we’re using… the way that Swan uses Swan is the same way.
So we use Swan to handle all of our customer touchpoints from the top of the funnel, creating awareness. To, converting that demand, to, having inbound meetings, pipeline, sales, post-sales, success, and support.
And the reason why is that that is just one unified journey of that customer, and context should be shared across all these touchpoints, and eventually what you see is that orchestration is pretty similar. You know, there’s, like, that CRM that you want to make sure that you’re working with, you want to send emails.
You want to ensure that, you know, reps are getting alerts, or marketeer is getting alerts, or support is getting alerts, that still… the mechanics are very similar in that sense. And when you have a shared layer that everybody can work on. it makes it very, very powerful and much more efficient, right?
Because everyone is aligned on what’s happening with the customer.
To your point about the fact that we see, again, like, fragmentation of thousands of different solutions out there, it’s just that we’re in the interim again, we’re at this period of time where nobody knows what’s the right answer, so everybody has, like, their own, you know, saying, and say, yeah, maybe we got it, maybe we got it, maybe we got it, and buyers are just lost in this solution, because nobody really got it, right?
And… you know, I’m biased, but I do think that we might got it, and I do think that when you’re working with a coding agent. then it suddenly… it clicks, and it’s super easy to see that it can do all of that stuff. It’s a coding agent. It can rewrite… if it can do something, it will write the code to do it for you, right?
You’ve seen the power of Cloud Code. If you experience it, you understand that it is extensible. And so, just a final note on that, Julia, I don’t think that the future is that, like, everyone will vibe code all of these solutions.
But if you have a dedicated coding agent for GTM, then you get the power of actually tapping into all of these capabilities just from that coding agent, because it was already designed. for GTM work. So you have a coding agent for GTM, and a coding agent for finance, and a coding agent for design, and a coding agent for code.
And by the way, Anthropic is building all these coding agents, right? These are different coding agents.
They were the one… the first ones to see it, because they built a coding agent for code, a coding agent for knowledgeg, a coding agent for design, a coding agent for cyber they’re building now, and then maybe next they will hit a coding agent for finance. They’re doing it already.
We’re gonna… we’re just telling them, go to market, that’s our niche, so stay out of it.

Julia Nimchinski:
Awesome, and let’s actually double-click on this one. So the first concern of every report you would read is, how do you actually stay defensible In the age of AI. I’m curious your thoughts here. So, you are doubling, three of you, become six of you, and what else? That’s it? That’s it. That’s it.

Amos Bar Joseph:
Yeah, so where’s the moat? So I think that… in the next 2-3 years, it will be super clear that we’re gonna have two types of, software businesses. coding agents. Where they are building for a human, okay? That’s the… they’re building for a real human user at the other side of it, trying to solve a problem for them. Is it design? Is it code?
Is it cyber? Is it GTM? Right? And that will be very little. you know, solutions, because there will be huge platforms who will dominate, like, their space. And then, the other component of software businesses will be Software that goes to the coding agents. Okay, so kind of like infrastructure for coding agents.
What we’re already seeing is that Salesforce is going headless, RAM built, you know, their entire business as a CLI, we’re already seeing all these businesses starting to reconstruct themselves as developer tools for coding agents, right?
So the ecosystem is already forming around coding agents as the main interface for humans and the main destination and platform. And so, I believe that there is a place for a coding agent for in each category. You can build it. The models. we’re not fighting, like, we… Anthropic and us, we get the same model.
We’re building the same coding agent on, like, on the same model, so they don’t have any intrinsic… you know, advantage over us because they are a model company, like, we… we’re both building a coding agent. And we have a lot of experience in the go-to-market, and so we’re building a coding agent with an amazing taste. and GTM.
And maybe in our next session, Julie will break down what does it mean, and the importance of taste in a world where, you know, software is super easy to build, and all is left is that expertise in problem solving.

Julia Nimchinski:
Love it. Amos, how do our community just test drive Swan AI? Where do we go? How do you do it?

Amos Bar Joseph:
Yeah, so I go to GetSwan.com, and we have a free trial, and 500 credits. Julia, we could actually think of giving something, you know, additional for the community members, if they would like to play even more. And, you just start chatting with Swan. Swan comes with a lot of context on how you work already.
It’s a beautiful conversation, you’ll have fun with it. spicy. This one is not, like, a boring AI like a lot of the other ones, so, I’m sure you’ll have fun.

Julia Nimchinski:
Thank you so much. Always amazing featuring you, and yeah, see you soon!

Amos Bar Joseph:
Alright guys, thank you so much.