How Agents Learn, Remember, and Improve Over Time, with João (Joe) Moura, CEO of CrewAI

The models that are kind of like not the top quality are getting cheaper and cheaper, but the ones that are top quality, they're getting more expensive. So people think about there's a runway to the bottom on the LLM price, but it's not true. The models that are the cream of the crop, they are getting more expensive over time. I think with GPT-5, we saw almost 100% increase, so they doubled it. And even with Claude 4.7, they actually changed the tokenizer and that already expanded in 35 to 40% the actual cost of running the model. So in the end of the day, unless we keep seeing and we keep seeing investments into open source models, I think that will create two versions of the future, right? One where more people have access to this intelligence and one where people have less access to this intelligence. And I think that will define a little bit of how much of power will be controlled by the few and how much power will be controlled by many in this future.

Welcome to the Main Branch podcast. This is your host, Cagla Kaymaz. Main Branch is a show where I talk to founders and early adopters about building, deploying, and scaling world-class AI startups. Welcome, Joe [Moura]. You're the founder and CEO of CrewAI, which has become one of the most popular frameworks in building agentic systems. Your open source project has over 50,000 stars. You've worked with over 50% of Fortune 500. I think last we chatted, it was close to 65%, which is incredible. I'm really excited to have this conversation with you.

Thank you so much for having me, and yeah, I'm very excited to chat about it. I got to say, running this company has been quite a show. It's just like one of those times in the industry where things are moving so damn fast that it is very exciting. But thank you so much for having me. I'm excited about this.

So assume I know nothing. Assume we have never chatted. Tell me about CrewAI. What's your one-minute elevator pitch?

Yes. So I would say CrewAI exists to basically help people adopt AI agents faster, especially companies, especially enterprises. So what do we notice is many of these companies are stuck, but there is a huge value on the other side if they manage to adopt AI agents. And a lot of that goes back to the fact that it is a very small subset of people that are able to deploy, build, and trust these use cases. So we have them close that gap.

As I mentioned, you're one of the most popular frameworks. These days, people are not calling a lot of things frameworks. It sounds like almost frameworks are dead and now we are on to harnesses. And with harnesses, you have other things like memory, reasoning, orchestration kind of built in. What's your view on these two concepts? How do they relate? How are they different? Are frameworks dead?

It's funny because there's a lot of conversation about, oh, frameworks are dead, frameworks are gone, harness is the new thing. And at the end of the day, I had a hard time wrapping my head around it, and take this as someone that is in the frontier of the industry. We're building these things. And I later made my peace with that because I was having a hard time because when I looked at what people have been calling harnesses and the difference between a framework and a harness, I realized that we have been a harness all along. The way that I would explain it for people that have no clue about what we're talking about is framework, think about the building blocks. So it's kind of like it's giving you a way to create an agent, giving you a way to do memory, it's giving you a way to do knowledge, it's giving you a way to do all these different pieces.

But you are the one that is actually stringing these things together, putting them together, making them work. And then it's almost like you have a spectrum. There's not a clean cut, but from that you start to get more and more opinionated all the way to the point that you have something like CloudCo. CloudCo just ... It is an agent, but it works outside the box. It already has memory. It already has knowledge. It already has all these things. So you don't do stitching things together. You don't even do any configuration. You just turn it on and it's ready to go. The more you get into the opinionated range, the more people start calling this a harness. So for CrewAI, it's funny because we started with a very opinionated, what some people would call back in the day of a higher level abstraction of agents because you create an agent that already comes with memory, already comes with this idea of delegation, already comes with this idea of learning and all that and just do some configuration.

And eventually we open sourced the framework behind it as well so people could actually use it. So that has been funny. But yeah, it's very confusing out there nowadays. I think people are trying to understand what I'm using. I'm using a framework? I'm using a harness? What I'm building and everything in between. But I would say in the end of the day, it's a matter of how opinionated that abstraction are. And the more opinionated, the more it starts to resemble a harness; the less opinionated, the more it's just loose pieces that you actually string along yourself.

Are you seeing any differences between certain type of engineers or certain type of companies that want to use a harness versus a framework?

Yeah, sure. It's definitely like, I think there's almost a place for both, in a way, where the framework gives you way more control. So in a framework you can control almost like if this then that in a way in certain ways, and you can have ways to swap memory one way or another and all that. And the agents end up being a little more out of the box so you can string them along together. In the reality, what we're seeing work most out there is both intertwined. So we see a lot of use cases where there are certain pieces that are going to be very deterministic and you want to have a lot of control on them, on how these things are going to actually work. But then you might have pockets of agency that you want these agents to just go out there and do all the work that you need and you want these agents to be pre-configured for you, and then people would go for that. That's kind of like the most common pattern that we're seeing.

When talking about, well, are frameworks dead, the case to be made there around frameworks being dead is this idea that you can build anything now. And like, well, do you need a framework if you can just prompt your custom framework into existence? And I don't see a world where that's real for frameworks and that is not real for harnesses as well. Again, harnesses is just like data framework with opinions on top of it, like things predefined. So I don't see a reason why that wouldn't be the next victim of LLMs in that way if you truly believe that those things are being killed.

If I heard you correctly, you're saying the harnesses could be the next victim or frameworks could be the next victim?

I think honestly, harnesses ... And you could build a use case that any software is the next victim. And in the end of the day, I think a lot of the companies are fearing that. The whole SaaSpocalypse story is about that. It's about people prompting their way into software instead of buying software out of the box and everyone trying to understand, well, how do I position my software on this? And I think that's a litle bit of what's getting people scared on this.

I mean, if harnesses are getting close to the chopping block ... I think you recently wrote about this entangled software and what comes potentially after harnesses. There is the entangled agentic systems concept. Maybe we can talk a little bit about that. What's your vision on what's the next big thing or next couple big things?

Yeah, no, that's a great question. I mean, a lot of people ask me this question. I think people look at us and say like, all right, you folks called multi-agents back in 2023 [when] no one was talking about multi-agents. If anything, I actually got a lot of trouble to talk about multi-agents because people would say multi-agents will never be a thing because when an agent hallucinates, if you do three agents, they're going to hallucinate elevated to three so there's going to be 95% hallucination. But now everyone agrees like, no, multi-agents is the way to go.

So people now come to me and ask, like, "All right, multi-agents are good. What is coming next?" The way that I think about it, it feels like where the puck's going is this idea of the agents improving over time. Because it doesn't really matter how much opinions you put into an agent, they still are leaving all those learnings from the actual executions on the table. So the idea of self-improving and self-learning, not only by creating memories, but creating its own skills, by creating its own flows, and then self-evolving and self-modifying itself seems to be a lot of where you can really get long-term benefits. And then the more I started thinking about this idea, I was like, well, these almost use another way to get into this quote-unquote same result that you would get by vibe coding something, right? If people think, well, I can create my own agent, I can create my own software, I can create my own SaaS instead of buying it and I would just vibe code my way there, I'm going to prompt my way there.

I almost see a reverse path to that where you get an existing agent or framework or harness or software and the more you use it, the more it starts to adapt itself to you. I think some people might be calling it adaptive software. I started calling it entangled software, entangled agents, and it's this idea of the agents learning from the usage and then even modifying themselves along the process. And then we start basically playing around with these concepts internally ourselves and get a few customers to try it out, and honestly, the results are just incredible. So I now have a strong belief that the whole idea about not only memory itself, but this idea of self-improving and self-learning is going to be the next biggest thing in terms of the cutting edge of AI agent development.

I'm fascinated by adaptive software. I don't know why my version of Affinity needs to look like your version of Affinity for completely different use cases, right? But it also feels like we're a little bit far out from that. I don't know if you're going to get there in six months, nine months, in a couple of years. Are you guys already experimenting with these self-learning agents? Do you have anything in practice you're already using?

Yes, we are. We actually have a few internal agents that are fully turned into that. So basically they are not only being used, but they're also always in watch mode. They're learning from conversations they're observing, conversations that they engage [in], conversations that they don't engage [in] at all, but they're just learning and capturing all those learnings and we're trying many different things. We tried the dreaming approach where every few hours they go over a dreaming state where they go over all their memories and then collapse them into learnings and all that. We did the self-creating skills, the self-creating flows route as well, that has been interesting. And for some of the more complex agents, we even make them open PRs against themselves so the engineers can then go in and can merge and they are updating their own code based on their learnings. And we are seeing massive results from that.

It's very interesting to see, if you go week-in, week-out, for a few weeks, things start just to work and then starts to have more integrations because it starts building its own integrations. It starts to understand more about your business units, your leaders, your teams and how they work together. And that has been honestly mesmerizing to watch sometimes. So I'm very excited to think about what the future of having agentic organizations actually looks like. We're now thinking about potentially creating ... We have this pod structure on our teams, on the engineering side. So it's all small teams. We're thinking about creating a full agentic pod where it's a product, a designer, an engineer, eveything is agentic in there, and see how that would work out as a next experiment now.

That's fascinating. I think you mentioned memory when you were talking about this, like using memory and getting learnings out of this. I think you also have a very differentiated approach to memory, like memory as cognition versus memory as storage. A lot of times when people think about agentic memory, they're thinking, hey, we're just going to store embeddings in a very fast vector DB, you're just going to retrieve information on the go versus, I think, you're using LanceDB under the hood, which I'm a big fan of, but you're doing something more sophisticated on top that includes importance, recency, and other metrics. What's your approach to memory today?

I'm so glad that you bring this up. I mean, you definitely did your homework going to the blog post there. By seeing the blog post, you can see how we are thinking about things and how I'm spending my time thinking about the technical side of things. So we started with this idea of self-improvement of like, well, for us to do self-improvement, we need to have a better memory. And that's kind of like how we start building the memory first and where that blog post came about. In the end of the day, right, let's just be very raw and real. Let's remove all the idea of harness, all the idea framework, let's remove all that and just talk about what's happening in there. LLM is intelligence for hire. So you're paying for getting some sort of intelligence, it's something that we were never able to do before.

So you're getting this token, right? And your job as an engineer in creating an agent is like, how do I make sure that the value for every token that I'm getting is more than what I'm paying, right? So whatever I'm paying for the token, I need to get that amazing value back. In the end of the day, the way that you extract the best token is by doing context engineering. That's the most simplistic way to think about it. So you're like, hey, I'm doing all this idea of put the right prompts, the right memory, the right things to create this context that when I send this message, I know that I'm going to extract the best intelligence, the best token that I can, and that token will basically be 2X more valuable for my company than I'm paying for. And memory goes a long way to that.

So a lot of the context engineering goes back to being able to provide the right amount of information. And if you think about a memory as just to store and retrieve data, you're leaving a lot of value on the table because the same way that we humans, when we think about something and you want to remember something, we just don't remember it. We create all sorts of associations. We start to remember how this is associated with something else and all that. A good example that we use is, for example, we at the company, we used to use one database provider and we moved to another database provider. If all that you did was just storing this into a database, now you have a database that tells you that you use two databases, right? But if you actually use some sort of cognition to, at the time of storing this, looking for similar things and kind of like trying to think about how you consolidate them, you now can store that you actually migrated databases, you can store why you migrated databases, you can store your learnings from that migration.

So there's a bunch of knowledge that is implied and that you can infer from even the comparison itself. So that's kind of like what we implemented at CrewAI, and we even run benchmarks on it and it's actually pretty remarkable.

I mean, for the example you gave in terms of which database you're using now versus what you used to use, obviously it doesn't depend on the intelligence under the hood, it's the factual information that needs to be stored as such. Do you find memory changing depending on the intelligence you're using under the hood? Like how strong and how powerful the model is versus like ... I mean, I want to also ask you about these token costs and open source models and so forth, or do you think memory stays the same no matter what the underlying model is?

I mean, honestly, you can remain the same, but you're going to be leaving some money on the table. You can definitely optimize it. So when you think about the bigger the model, the more expensive it is, the more smart it is. So that means that you can basically extract more knowledge from it. The other thing is that, depending on the model, like if it's a different family of models, that probably means that a different prompt style needs to go off that. So like not exactly the same context engineering that you would do from Tropic is the same one that you're going to do for kind of like a GPT 5.5. So what I think about that is yes, there needs to be changes in there. I would say it's less about the memory, though. It might be more about how you leverage that memory.

The way that it works in CrewAI now is when an agent's trying to store an information, to store a memory, it actually goes through that process that I described from consolidated, from understanding importance, relevance, and all that and saving that in the database. And some of this metadata actually decay over time. So it almost starts forgetting over time like if it's not relevant anymore. But then the process of retrieving the data also is an agentic system where, depending on the model that you're using and how much context you need, it might not only retrieve that information but actually pull the entire thread of related context that you want to inject in the model. And you do want to optimize that per model. You don't necessarily need to and it will just work, especially in the bigger ones, but the more you get into particular models, the more you optimize the better.

Again, at the end of the day, folks that are doing the agentic engineering, their job is [to] get the most valuable token out.

I mean, this is I think a good segue to talk about these token costs, right? You work with a lot of enterprises that have very scaled usage. I'm assuming they're starting to worry about their token costs if they haven't already. What are you seeing, especially for larger Fortune 500s in terms of their adoption of maybe open source models or interest in trying something local?

I mean, you go back a year, right? I would tell everyone that I thought that fine-tuning would be huge and all that. And it's funny to see that that didn't come true, right? Fine-tuning is not as huge as I thought it could, even though I think that it could actually be extremely big. But then I didn't even see many people using local models. A lot of people are just defaulting to either the hyperscalers or kind of like the big labs, and the hyperscalers just because they're throwing credits around so much, right? If you're running a customer from Azure and you have this huge commitment to spend $10 million with them, well, you better use it on tokens, right? So people were using it in that way, but I think now I'm starting to see people actually using local models. So I can't name the customer, but it's a huge customer on the FSI branch and they are actually using a bunch of open source models now.

They're [setting] up things like Qwen, DeepSeek, Qiyamah. There's a bunch of new models coming up out of Nvidia now as well, new Nemotron models that are actually pretty damn good. And even new models like Gemma 4 and now Google released under a patch too that people are trying out and using it. So it's funny because everyone saw, I think, the Uber news a couple weeks ago or three weeks ago where they're saying like, "Hey, we had a budget of tokens for the year and now it's like May and we're already done. We're already using it before."

And we're seeing more and more companies actually come up with that. So I think there's two things that are going to come from this. There's going to be a chair dance a little bit where people are going to be a little more strict with their token consumption. So you got to win your right to spend those tokens. So for companies like us, that means that we need to step up big time on ROI, show you the value and all that. I think people that have been selling on based off FOMO, they're going to see a lot of that start melting away for them. And I think that the other thing as well is there's going to be more people open to not only fine-tune, but local models in general. So I think that's going to be a trend this year.

The beauty with Crew is it doesn't matter which model they're using, right? They can still work with you. I think you're safe on that front, and I'm with you on the fine-tuning. I remember thinking two or three years ago, enterprises have all this data. They are in the best position to fine-tune or even train eventually, but we haven't seen that taking off. I think this general intelligence has been good enough so far. I think people are still okay overspending potentially and still figuring out the ROI part, but there's so much pressure for AI adoption that I don't think people have started to optimize using fine-tuning or other methods as much. So you work with a lot of enterprises, as we just mentioned, and what's fascinating is not all of them are tech companies, right? Fortune 500 has other companies like CPG, logistics, manufacturing, and I think you've worked with a bunch of them.

What kind of use cases are you seeing them adopting? Let's take out the tech companies in the Valley. What are you seeing outside of the Valley in terms of enterprise use cases?

Yeah. I got to say, funny enough, that's where we're seeing most success as a company. So if you look at our customers, our customers are not your Ubers and DoorDashes, right? Our customers are like Johnson & Johnson, PepsiCo, AB InBev, Experian. It's these large enterprises that they do have amazing technical teams, but they're not necessarily associated with like, oh, this is like a tech company, the same way that Uber is. So that's kind of like where we're seeing a lot of our success in there. And honestly, it's fun because if you think about it, with all this conversation about AI, you look at even like an S&P 500, and you see the Mach 7 just skyrocketing. But what about the other 493 companies? A lot of the companies in there are not tech companies and they're the ones that are promised to get the value of AI, right?

All the investment and infrastructure is based on this promise that like, hey, you folks, you get the value from this. And they're trying to understand how to grab this bull by the horns. I think we are showing up with a platform that is complete, there is batteries included, with experience knowing what's working, and they're choosing us for that. If I think about the use cases, it's interesting because there's almost like a maturity curve. So when they initially come to us, they come [with] a lot of thinking about like — how do I put it — like savings, right? They want to save money, in a way. So they're thinking about, well, how to reduce costs with AI agents. But then right after that, they start thinking about, all right, how [do] I generate revenue with AI agents? And then right after that, they start thinking about how [do] I innovate, like what are things that I would not be able to do?

We have a company that is a huge CPG, and they started with a very simple use case. It was kind of like a discount use case. So they basically had a dedicated team to do a bunch of approvals for discounts and they basically replaced all that with CrewAI and they moved that people into something else, and that was massive for them. We're talking about 96-97% efficiency gains in terms of how many hours they're spending on this process.

Help me understand the discount use case.

They have around 150,000 employees and they have a policy that many teams can just ask for discounts. And you would have [the] marketing team asking for discounts for doing events, you would have sales teams asking for discounts to help closing a customer, whatever it might be. And that was a very manual process for people actually approving. There was parts of that process that was very deterministic, but there was parts of the process that had a judgment call or like someone would to say, "This is approved, this is not approved." And they end up ... That volume of those requests became so large on a global level that they had a dedicated team that all their job was just approving or denying those processes. And then they started experimenting with agents using CrewAI to do that process and they basically built the agents and they did a backwards test.

So they run like last 30 days, how much the agent would have approved or denied, and they found that it actually matched the human behavior 94% of the time. And they're like, "This is really good. If we can identify the other 6% and we can delegate that to a human, that feels like we could grow this out." And then when they roll it out, they realize that the amount of time they're spending is now like only 3% of what they used to spend before. Because now when people show up to do something manual off one of those 6%, it's more of like everything is in there, they have a description, they have an understanding, and the bulk of the requests are already automatically approved. So that was their first use case. It was very kind of like efficiency gains, cost-saving kind of use case, right?

But then you look at some of the latest use cases they're doing, they have been a customer for over a year now and they're doing use cases where they're using agents to automatically approve legal documents and things that they just couldn't do at scale. I won't name the customer but I'm going to tell you this: They have hundreds of beverage brands under them. So there's only a couple of companies that would fit into that. They had this thing where they would go into their customers or potential customers and they would say, Hey, I'll get you the paperwork that you need to have my product and I will get you that paperwork, figure it out. And that will cost me a million dollars, that would cost me a lot of time. But if I do it, then you buy for me for their first 5 million of purchase.

Those customers would say like, "Yes, for sure." I mean, "I don't want to figure that out, I don't want to go through that process. Let's do that." But the problem is that that worked so well that they now have thousands of those licenses that they own, but it's not theirs, it's for their customers, and they need to renew that process every year and they now lose time, they get fines, it becomes a nightmare. So what they decided to do is basically use agents to automate the whole process of those license renewals. And by doing that, they now, yes, they're saving money because they're not getting millions that they're expending before between fines and people on this, but they actually get to expand this growth hack across the globe. So they're very excited about that. So again, it is funny to see how, usually, they start with like, "I want to save costs, I want to get efficiency gains." But eventually they start to think about, "Well, I couldn't do this at this scale, but now I can with agents." And then they start going [on] a better route.

I mean, those use cases are so different from each other. At the same time, your platform is fairly flexible to serve both of those. I'm assuming for some of these customers, you go in, they've not built an agent before, they're like, "Where do we even start?" I remember watching your deep learning course a couple of years ago, you were pretty good at bringing people along and teaching them how to use these systems. One thing I want to try with you today, if you're open to this, I want to see how you would workshop something like that. You go into a Fortune 500, you work with a bunch of senior engineers, they have some sort of a workflow they want to automate in their mind, and they're saying, "Hey, Joe, help us architect this. What do we do?" I would love for you to walk me through that.

Let me do this. I have a easy one, but again, the sky is the limit. So let me share my screen. In order to keep this simple, I'm using the platform alone. The open source will allow you to basically code so you can do a lot of super complex things in there, but I think this will make it a little easier for us to understand. Usually when these customers come to us, if they already have a use case, there's a process of understanding the actual use case. If this use case is going to be more deterministic, if it's going to be less deterministic, and what that's going to look like. So we do go through a process of actually sketching it out, trying and thinking and all that. If they don't have a use case, we actually have this idea of discovery. Discovery allows them to understand what they should build.

So I can come in here, for example, and let me say Coke. Coke is not a customer of ours, but if they were, I could say like, "Hey, Coke wants to build some use cases." They have BigQuery, they have Databricks, DocuSign, they are a huge Microsoft customer, and they use Zoom, whatever it might be, right?

And Joe, this would be their teams using it, or this would be your team or you walking them through this workflow?

That could be both. So at this stage, that could be their team, but we have a deeper access on our side that allows us to do more fancy stuff.

There's a few building blocks in here that people might care about. One is agents. You want to use as many agents as you can. So you can create agents. We have some in here that we use very often, but you can create new ones or you can bring external agents as well. So we have companies right now that have CrewAI managing, for example, ServiceNow agents where you can bring ServiceNow agents using A2A or even like Agentforce agents or ADK, LangGraph agents, whatever it might be. As long as it's a part of A2A, you can bring them in here. The other big building block is this idea of tools, right? Things that your agents can use. We have a bunch built out of the box so you can use all these different connections. You can bring your own MCP servers if you want to, or you can create custom ones. So you can see that we don't have an integration with Phantom so I ended up creating my own and people can actually use this in our organization. I control who has access and everything.

The way that I like to use this is basically by just chatting my way into an automation. I have one in here that I got already just for us to try it out. It's kind of like a simple one, but I could deploy this in any sales organization and it would work. So this says like, Hey, given someone's email, I want you to research them and their company deeply and then cross any relevant data in our HubSpot with that and then cross any emails that I have exchanged in my Gmail with them and prepare me for a sales call.

But I'm sure you get this semi-annoying question often, someone comes to you and they say, "Hey, actually my salesperson, they just created the skill in Claude Code. They have this research agent they're working with like, why do I need Crew? Why don't we just do this in Claude?" What do you tell them?

Yeah, I got to say, the cool thing is Claude has proven to be a great tool for us, actually, because what we're doing is we connect the platform directly in Claude. So we understand that people don't necessarily want to come into the platform, and I think that's fine. I mean, you got to meet these customers where they are. What we're finding is that more than Claude Cowork, like bringing these things into actually IEMs like Slack or Teams is where people are really getting to usage, but we actually have an integration straight with things like Claude Cowork. So you can in there run the same use cases or pull the same things, but now you go from more single player like Cowork to something more multiplayer like Crew. And on top of that, you start to have all these settings where you can do not only quality and visibility, but add humans [to] the loop, cost alerts, remove PII.

You have controls around your expenditures, your use cases, how many tokens are going into LLMs and all that. So it's more kind of like enterprise-ready that gives availability for that like InfoSec and IT team to feel comfortable and to actually deploy these things and allowing people to use them.

This is pretty cool. And it also gives us a sense of what do you have outside of open source, right? A lot of times, I'm a big fan of open source, but sometimes founders give away too much into open source and then it's hard to monetize, commercialize it after. How did you approach it? When your open source project is extremely popular and then you have the Studio and you have a couple other enterprise features that are not part of the open source. When you decide on what to put into open source and what to put behind a paywall, how do you go about it?

Honestly, I think there's ... There's two kinds of open source companies, I would say. There's probably more, but there's two on a high level. One is kind of like the Mongo playbook where you say like, hey, I have Mongo open source and I'll stop using Mongo open source and I'm going to start to use Mongo enterprise. So you're migrating. And then you have companies like Vercel where I'm using Next.js. I'm not going to stop using Next.js, but I'm going to bring Next.js into my Vercel account. And we are more of the latter. So it's not like, hey, the enterprise replaces the open source, but it's more like you get whatever you're doing on the open source and you bring it here.

Whatever you build on the open source, you can deploy here and you have all the same metrics and the same features and everything in between. So the idea is to extend the open source. Now we do, behind the scenes, have a private fork off the open source that is when you deploy is what we use and that is faster, supports load, and has a bunch of other things as well that we do behind the scenes that's exciting. But in the end of the day, you're not stopping using the open source for this. And the open source is very geared towards that engineering persona that is trying to build something on their computer. But the gap between that and getting something into production, the last mile is more like a thousand miles. There's a lot in there from the integrations, the access, the controls, and all that that do get in the way.

And in the end of the day, a lot of our deals and our customers is less about that and it's more about accelerating adoption as a whole. If they say, all right, I have a subset of people or a subset of engineers that can do agents that allows me to move fast. But if I can unlock many of builders, not necessarily even only engineers, but I can unlock many builders throughout my company and bring all the power and the ease of use of CrewAI into the people that are actually doing the work and they can automate it in a way that is basically complying to central rules, then that is where transformation really happens.

I agree with you. A lot of times there's so much more after the demo that needs to be done, but at the same time, open source is so helpful to just get started. I hear a lot of enterprises saying, "Hey, before we adopt, especially a dev tool, we want to try it out." You don't want to go through the whole procurement and onboarding. So it's much easier for them to try out the open source. Another thing I want to get your thoughts on, you've done such a good job building the community and the brand and the narrative around Crew. What are some things that worked for you? Let's say you were a DevTools founder starting today, you're building an open source, what are two-three things you would definitely do to build out the community and build out the narrative?

Well, it's funny because I always loved open source. Before CrewAI, I had other three or four projects that were what people would call successful. Nothing as big as CrewAI, but things that got millions of downloads in a month. At the end of the day, it was all about the fact that you got to be using it. I think the reason why CrewAI grew up to be what it is from the open source side was actually because I needed that. I needed something and I did it for myself. I never set out to say, "I'm going to launch a framework. I'm going to launch a harness. I'm going to launch a company." I set out to say, "Hey, I'm building these agents and damn, it doesn't feel like very economic." And it was like, I wish there was something out there that was easier to use and meant for production.

And that's how I started to abstract those things away. And I just started to put myself out there, and then the more that I put myself out there and the more that I talked about the framework, I think more people start to resonate with those ideas.

How did you put yourself out there? What are some concrete things you did?

I did a lot of video recording. I mean, I remember my wife would get pissed off at me because she was like, "Hey, we need to leave in half an hour." And I was like, "Let me record a quick video. I didn't have video today still." And she knew that once I start recording the video, it's going to be like two hours. But I remember getting late into a bunch of things that we would commit to just because I was recording videos and then editing them on the way. She was driving and all I had was my computer just editing, adding subtitles, and everything. But it was just a lot of doing that, just like doing tech talks, pushing videos. And again, it didn't feel like marketing. It felt like here's the cool thing that I did. Here's the cool thing that I built, a new agent, a new crew, something that I'm using for myself, and that really resonated with people back then.

The other thing that I would say if I had to do an advice for someone is like, be opinionated. I think another big thing about why CrewAI worked out is we were very opinionated from the get go.

So be a harness, not a framework?

Not necessarily be a ... Yes, be a harness in the context of Crew of ... No, but I think you're right. It's like being a harness on the context of agent, but even on the framework level, if you say, "I'm going to build a framework," be opinionated about what you're giving it to users.

Think when people feel like — especially in a new tech like this where no one knows how necessarily to be — there's no right or wrong, everyone's trying to figure out what is the conventions, what are the new protocols and everything. There's something to be gained by saying like, "This is the way that I think I should do it." And if you're wrong, no one cares, no one will bat an eye, but if you're right and people appreciate that, then that is the opportunity where you can take off. So I think from the get-go, that's how and why I say we first built a harness before we built a framework. Out of the gate, we came with very strong opinions of how these agents would work and how they would be and the fact that it was multi-agents and the fact that they all would be roleplay, and people really resonated with those opinions.

And how do you think about today? Do you have dev rails or are you still doing a lot of that yourself?

I think the fact that I did that for so long was detrimental for the company. And the reason why I say that is it took me a long time to start delegating that and I think that almost like that created a culture that a lot kept coming back to me on the marketing side of things. Now we're changing that and I'm happy about it, but it took us extra effort because for a long, long time I was the one that was out there doing videos every week and putting things out there. I still want to do it. It's just that I'm spread super thin nowadays and that makes it harder.

You are the face of the company, and it sounds like you enjoy the videos and the tech talks and all of that. So I'm sure part of it is you wanting to continue to do it, but you only have so many hours in the day and you need to delegate a little bit of that.

Exactly.

I have two quick questions for you to wrap this up. What's the vision for CrewAI? No one knows what's going to happen in five to 10 years, but if you had to guess, what is Crew doing in five years?

There's companies that have a small TAM, there's companies that have big TAM, and then there's AI agents, right? (laugh) When you think about AI agents, we're talking about the future of work. So I think that's one of the things that makes it so hard to predict where things will go. But I can tell you a little bit of what I want for the company and what I want for our team. We're a company that we're very empathetic as a company. I think we understand that we put ourselves in our customer's shoes. I think we understand why AI agents are so valuable and how transformational they can be, but we also understand that a bunch of these companies out there, they're struggling. You go into Twitter, you go into X and LinkedIn and there's going to be people celebrating, celebrating, celebrating, but you go out of that bubble and you go talk with these real businesses out there, they're struggling actually adopting the stack and getting this out there, and they're almost getting sick of hearing about tomorrow's new thing while they're trying to deploy yesterday's thing still.

So I think there's a lot of helping them getting there and build a product for them to help them getting the adoption. So I would love for us to grow into the kind of company that enables these companies to become agentic, to materialize that future for them. And you're going to have people that are responsible for managing like not only this multitude of agents within their business unit and their organization, but you're going to have those people within these companies that are responsible for controlling the guardrails and all the rules and policies and everything. So I think that's a little bit on the future that we're going, and even us internally, we are seeing that happen, right? There's people inside CrewAI, they're becoming more of like this control plane managers. They're keeping tabs on what is going on and helping make sure it's doing away while enabling everyone on the edges to actually build their use cases and start distributing them.

So for me, it's very clear that, one, the genie is not getting [out of] the bottle. There's no world where people are going to stop using AI. There's no world where people are not going to use AI agents and all that. But definitely right now, the budgets are being way more scrutinized. People are way more focused on these things like actually seeing the light of the day is less about volume of use cases and more about actual outcomes and results that you're getting from these use cases. So it's almost like the tide is changing in terms of now everyone wants to get real and we're about to see, all right, who is actually real about this and not? So I think that's going to be an interesting point in time.

Yeah. I mean, I think about this gap between AI capability and adoption quite a bit. I think that gap right now is increasing, unfortunately, but I think we're going to hit a point where the gap starts, hopefully, at least keeping up and not increasing it more. Because when you think about the bubble we live in versus some of the outside of the Valley companies or even internationally, right? We're just scratching the surface of how people are using AI agents and there's so much to be done, and it sounds like CrewAI is going to continue to play a big role in this.

My last question for you is, this is an even harder one because we're going to go way to 2050, like stretching your imagination. What's happening in 2050? No one knows, but are we living in this post-scarcity Star Trek universe where things are peaceful and we have these AI collaborators working with us? Or is it something else? What's your vision?

I think a lot hinges on a few things. I think we will go back to open source models, I think so. I think for me that we're running a company, we're using AI so much and LLMs so much, it's easy for you to forget that there's almost like two prices going on. The models that are kind of like not the top quality are getting cheaper and cheaper, but the ones that are top quality, they're getting more expensive. So people think about there's a runway to the bottom on the LLM price, but it's not true. The models that are the cream of the crop, they are getting more expensive over time. I think with GPT-5, we saw almost 100% increase, so they doubled it. And even with Claude 4.7, they actually changed the tokenizer and that already expanded in 35 to 40% the actual cost of running the model.

So in the end of the day, unless we keep seeing and we keep seeing investments into open source models, I think that will create two versions of the future, right? One where more people have access to this intelligence and one where people have less access to this intelligence. And I think that will define a little bit of how much of power will be controlled by the few and how much power will be controlled by many in this future. I truly believe that there's a lot of incentives for people to actually keep pushing open source models. I know there are companies that are deeply committed with that. I'm very excited about what Nvidia is cooking, for example, with Nemotron models recently, and I think that is great. We also see other great models coming from outside of the US, right? We're seeing the Deep Six, the Kimis, the Qwens, we're also seeing the Mistrals and all that. There is kind of like a race on that open source model stage as well. And as long as those keep getting better in a similar rate to the closed-search ones, even if they're not at the same level just yet, I do feel more optimistic about the future, but it's very clear that the future will be very Star Trek-y in that way. It's going back 50 years ago, even five years ago, you wouldn't believe that it would be working the way that we do. For engineers specifically, they I think are feeling it the most. You go back six months ago and people would not have guessed they would be working in the way that we're doing. So yeah, 50 years from now, I would bet on robots, physical AI, physical agents, and everything in between.

Lots to hopefully look forward to. This is awesome, Joe. I know we are at time. Really appreciate you joining me. I had a lot of fun.

Thank you so much. Thank you for having me. This was great and I hope everyone watching and listening in had a lot of fun.