# The rise of AI agents

- **Show:** Gradient Dissent
- **Published:** 25 Feb 2025
- **Source:** https://wandb.ai/site/resources/podcast
- **Transcript:** https://podcasts.opennash.com/shows/gradient-dissent/the-rise-of-ai-agents-765d29/
- **Provider:** OpenNash Podcasts (machine-generated)

## In brief

João Moura, CEO of CrewAI, argues that large companies will soon run thousands of AI agents. In his view the hard part is managing them across their whole life: planning, building, deploying, monitoring and improving them. He says real agents are defined by agency, meaning the AI decides what happens next. He also says the most useful agents connect to a company's internal data and systems.

The episode contrasts rapid adoption with real limits. Moura reports heavy usage and big-company customers, yet says it is early days, and most work still has humans in the loop. Agents excel at problems with clear right answers but struggle with nuanced business judgment. He expects the next step to be agents helping with decisions, which also puts pressure on software pricing models built around human seats.

## Key insights

1. **Real agents choose their own next step** Moura says an AI workflow becomes an agent only when the AI controls what happens next. A retrieval app that always runs the same lookup and summary steps is, in his view, just 'if this, then that', and he calls relabeling such tools 'agent washing'.
2. **Agents gain value from internal company data** Simple research and scraping tools cover many easy cases. Moura says the real value comes from connecting to systems like Salesforce, SAP and long-running homegrown software, through an API or a data lake.
3. **Control comes from narrowing each agent's scope** For a 60-page IRS form with a 720-page instruction manual, regular code splits the form into questions. Small groups of agents then answer one question at a time, which limits hallucination, and humans can be inserted at any step.
4. **Agent seats may undermine software pricing** Moura says per-seat pricing is awkward for agents. A cheaper agent seat could cut a vendor's revenue, and a very good agent might remove the need for a seat at all. He expects agent-specific endpoints to arrive later than most people think.
5. **Agents are strong at math but weak at nuance** Moura says reasoning models excel where answers are clearly right or wrong, because reinforcement learning can train on that. Business decisions are murkier, so humans still oversee most agent work for now.

## Questions this episode answers

### What is the difference between an AI agent and a regular automated workflow?

Moura says a workflow follows fixed steps, even if AI helps along the way. It becomes an agent when the AI itself decides what happens next, such as choosing between options A, B or C.

### Which kinds of companies are actually succeeding with AI agents?

Moura says successful ones have active leadership support, a technical person who can connect internal systems, and a clear use case in mind. Companies that just ask what they should do with agents are usually a poor sign. He also named finance, insurance and GCP (Google Cloud Platform) as fast-moving sectors.

### How do you stop an AI agent from going off the rails?

Moura describes limits on the number of requests and on running time, such as a 60-second cutoff that forces a best answer. You can also add programmatic guardrails, meaning code that checks an agent's output before it is accepted, and you can force a human into the loop.

### Will AI agents take people's jobs?

Moura says he doesn't think we're close to that. He says agents are mostly absorbing repetitive work: where four people did a task, one person now oversees agents and the other three move to more interesting work.

### How might AI agents change how software companies charge?

Moura says charging per human seat gets tricky when agents are the users, since a very good agent could replace the seat. He says LinkedIn is experimenting, and expects pricing to change as agents consume data instead of humans.

## Facts

- **João Moura** [0:12:30]: "I remember this person approaching me and saying, like, Hey, we are Oracle, we're using we're using CrewAI in production."
- **Lukas Biewald** [0:11:46]: "And now your website says that 40 of the Fortune 500 uses, um, CrewAI."
- **João Moura** [0:14:53]: "And January alone was, uh, over 50 million agents. So that was, yeah, that was insane."
- **João Moura** [0:15:32]: "I remember seeing at once a like a crew that had 21 agents on it, and that, for me, was was impressive."
- **João Moura** [0:03:51]: "when we see kind of like a media company using CrewAI agents to automatically edit files."
- **João Moura** [0:06:36]: "and now they have a lot of their pricing flows automated. So they adjust prices using agents on the runtime, uh, in marketplaces and other spots"
- **João Moura** [0:07:33]: "we're talking with some folks from, uh, Gartner, and I think they're referring to this as kind of like agent washing, like, um, companies that are stamping this is agentic now"
- **João Moura** [0:10:29]: "I mean, we're seeing some companies experimenting with that. I think LinkedIn is doing something, uh, in that in that lines."
- **João Moura** [0:19:25]: "it was a form that was around 60 pages long, but fear not, it comes with instruction manual, and the instruction has 720 pages long."
- **João Moura** [0:27:43]: "You can say like, Hey, you can only run for like 60 seconds. And then what happens is as it gets to that 60 seconds, we do a final call saying like, Hey, all right, you're done, your best answer right now."
- **João Moura** [0:23:40]: "That automatically becomes an API, and that's production graded. So I'm talking, uh, load balancer, auto scaling, SSL, like all those different things."
- **João Moura** [0:25:36]: "we do have all the agents in CrewAI, open source in an enterprise. They have short term memory, long term memory, and entity memory."
- **João Moura** [0:26:46]: "But that gets reset in every run. So you do a run and then boom, everything gets cleaned up and a new run happens."
- **João Moura** [0:26:46]: "Now, the long term memory is where they store learnings that are coming from multiple executions."
- **João Moura** [0:36:05]: "So it got to a point in time where we were overriding like 90 of the code that we were importing, and that was one specific piece, the class that we were importing."
- **João Moura** [0:37:52]: "I think they have like over 300 tools that you can also use with CrewAI."
- **João Moura** [0:29:21]: "finance and insurances, they usually want to self-host their models because they wanna run the whole thing in kind of like a airtight container"
- **João Moura** [0:33:13]: "what we're seeing these models like really excel, especially cognition models, is on the math problems is because they have a hard, right answer, right? So they can do reinforcement learning based on that"
- **João Moura** [0:45:03]: "we have a what I call a gradually evolving UI. So you start with a chat interface, similar to, uh, custom GPTs back in the day."
- **João Moura** [0:39:46]: "is that you can add custom cursor rules, and I think people don't realize that, and you can do that per project."
- **João Moura** [0:46:24]: "we have crews reviewing every single pull request across the company, open source and closed source."
- **João Moura** [0:46:53]: "So as these customers onboard the platform and they put their names and emails, we kick off agents that research them"
- **João Moura** [0:48:07]: "I think that's gonna be a trend this year where we're gonna have smaller models that are very, very capable, like better than like 70 billion parameters models nowadays."

## Transcript

**[0:00:02] Lukas Biewald:** You're listening to Gradient Dissent, a show about making machine learning work in the real world, and I'm your host, Lukas Biewald. All right, this is a conversation with João Moura, the CEO and co-founder of CrewAI, which is one of the leading AI agent platforms. He is on the forefront of agents and their applications to enterprise. So we talk about how agents work on the CrewAI platform. We talk about which companies are using agents successfully.

**[0:00:30] Lukas Biewald:** We talk about what tools those agents call most frequently, and we talk about how software companies are actually reshaping themselves for an agentic world. I hope you find this conversation useful and enjoy it. All right, Joe, thanks for taking the time. I really appreciate it. Um, can you start by describing the problem you're trying to solve with CrewAI?

**[0:00:52] João Moura:** Yes, uh, and by the way, thank you so much for having me. I'm very excited to be here today. Um, so the problem that we're trying to solve with CrewAI is that three years from now, most companies, especially enterprises, are gonna have thousands, if not hundreds of thousands, agents working for them. And either that's gonna be thousands or hundreds of thousands legacy code bases, or there's gonna be something they're gonna be able to manage in an actual asset.

**[0:01:23] João Moura:** And we are building the control plane that allows them to do that.

**[0:01:28] Lukas Biewald:** Interesting, so what do you expect these agents to be doing?

**[0:01:32] João Moura:** Ah, I think it's gonna be a little all over the place. I mean, we're already seeing that there's no clear winners within these companies. It's very much cross-horizontal. So there are people that are doing backouts automation, people that are doing coding automation, people that are doing support. So it's a little bit of everything. And if anything, that just makes the challenge a little more exciting to me. Just because if you if you prove like this is doable in one specific horizontal, it's very easy for you to now start talk about all the different horizontals that you can expand to.

**[0:02:03] João Moura:** So I think it's gonna start with what we call low precision use cases, where you still have humans in the loop, people are reviewing things, but I think it will gradually kind of like evolve into more decision making, and things are gonna get very interesting by then.

**[0:02:21] Lukas Biewald:** And what does the control plane do?

**[0:02:24] João Moura:** Hm. Well, when you think about the life cycles for these agents, right? They start with planning them. So a lot of people are focusing nowadays on beauty, but beauty is just one component, right? And there are there are frameworks out there. Gray AI is one of them, probably the best one, if you ask me and many other people out there. But it's not only the beauty. You got a plan, and then once you build, you got to deploy, and once you deploy, you want to monitor, and once you monitor, you want to iterate on this.

**[0:02:53] João Moura:** And that is not not even mentioning things like authentication, scoping, access, marketplaces, and everything in between. So what we're trying to do is across their entire like stack and make sure that these companies are equipped to build, deploy, monitor, and iterate those agents.

**[0:03:13] Lukas Biewald:** So what is working today? What are companies actually doing with agents right now?

**[0:03:18] João Moura:** Nice. Well, I gotta tell you, it's kind of impressive. If you look at, like, the the more simple use cases, there's a lot of people doing sales and marketing and all that. Then you have kind of like more advanced companies that are using some back office automations where they're automating like more custom processes within kind of like their their businesses. And then you start to getting into kind of like some of the cutting edges, and you find companies trying to tackle very complex problems, like automating entire life cycle, kind of like code creation, or trying to automatically fill r r s forms.

**[0:03:51] João Moura:** What is something that you don't want to get wrong? Uh, and then there's like some folks that are more into, even more on the cutting edge, when we see kind of like a media company using CrewAI agents to automatically edit files. So you have live footage of games streaming on TV and then agents that can track the ball automatically editing, cutting, adding captions and sound, and pushing that into social media. So there's a little bit of everything, but I would say it's very early days.

**[0:04:24] João Moura:** A lot of these companies are early on their journey and starting with kind of like simpler use cases that they can then scale on up.

**[0:04:32] Lukas Biewald:** What what kinds of companies are successful deploying agents, like, what are they doing? Because, you know, when I talk to most companies, they're interested in agents, like there couldn't be, like, a hotter topic than agents right now. But most of the companies that I talk to haven't yet successfully, um, gotten agents to do anything except maybe, like, a little bit of, um, coding, sometimes chat support, sometimes a little bit of, like, sort of internal support. That that's those are the places where I see things working.

**[0:05:00] Lukas Biewald:** But it seems like you're probably talking to companies that really are, like, nailing this. What what? Uh, why is that?

**[0:05:07] João Moura:** Well, I think there's a combination of things, right? And it's funny that you ask that because that automatically becomes some of our qualification criteria on deciding, how what companies we engage more deeply with, right? Because, as you said, it's a very hot topic right now. So everybody and their mother wants to talk about AI agents, and uh, because we're kind of like the leading platform, everyone comes to talk to us. So what we gotta do as a business is kind of like being able to filter out, like, who actually is real and check a bunch of those boxes so that we can go deep with them and really bear hug them.

**[0:05:40] João Moura:** So, usually the signals that we look for for the companies that translate into successful use cases is there is this active support. So there is this active. They're kind of like saying, like, hey, we really need to embrace this. We know this, like, where their business is going. There's usually technical support as well. So there's like, um, even if the buyer is not a technical persona, there's usually like someone technical within, kind of like the project and the scope that is gonna unlock those internal integrations with kind of like homegrown systems and whatnot.

**[0:06:11] João Moura:** That really unlocks kind of like the power for more custom use cases. And then there's just like an understanding of, like, what a use case is they're actually want to pursue, right? Like, when, like, for example, if a company approaches us asking what they should be doing and what we're seeing other people do, that is usually not a good signal. Because there's not necessarily like a clear pain or something that they're trying to do, and that they haven't spent too much time thinking about that.

**[0:06:36] João Moura:** So those would be some of the signals that we look at. But now, on the day-to-day, there's a lot of kind of like starting small and then expanding into other things. But we have, for example, uh, more advanced customers. There are like Fortune 500 companies that actually started with simple use cases, and now they have a lot of their pricing flows automated. So they adjust prices using agents on the runtime, uh, in marketplaces and other spots, kind of like agents monitoring competitors and a bunch of other things throughout their company.

**[0:07:10] João Moura:** So very interesting use cases that we're seeing out there.

**[0:07:14] Lukas Biewald:** Interesting? Um, I guess, do you have a, um, a specific way that you define agents? I feel like, uh, like everything in AI, once it gets hot, the sort of definition expands. Like, would you consider, like, a RAG application an agent, or does it have to be, um, more advanced than that?

**[0:07:33] João Moura:** I think it's not about how advanced it is, but I would not consider that. And it's funny that you say that because we're seeing, uh, I think we're talking with some folks from, uh, Gartner, and I think they're referring to this as kind of like agent washing, like, um, companies that are stamping this is agentic now and kind of like trying to pass by that. And I think that is, uh, that is in the short term detrimental for the industry, but in the long term, I think things are gonna things are gonna consolidate and gonna figure themselves out.

**[0:08:01] João Moura:** Now, my definition is that agents require agency, uh, so you can have workflows, and you can have AI help with those workflows. I think at the point where the AI is actually controlling the flow of what is happening next, I think by them, you have an agent, because the AI is actually guiding the process and they have the agents to choose between A, B or C.

**[0:08:28] Lukas Biewald:** So, like, a RAG application wouldn't count because it always uses the same tool for the lookup and then the same tool for summarization. It kind of just will, like, choose, choose its own path. Interesting. Okay,

**[0:08:38] João Moura:** exactly. It's kind of like a if this, then that, right? Like, if this, then that, like that, that is necessarily an agent. Now, if you get the right output and you do something with it, and depending on what you're doing, like, different things might appear, and you don't necessarily know what that might be, then you have an agent.

**[0:08:54] Lukas Biewald:** and now, what about, um, tool use? I mean, tools obviously are like a critical component for adoption of agents. What what kinds of tools do you see used most frequently, and and where is this going?

**[0:09:08] João Moura:** Yeah, I think tools are kind of like what make these agents very useful, like, near end of the day, right? You got to have, like, uh, tools that allows these agents to connect with kind of like internal data or external data or whatever you want. I think there's a lot of kind of like the simpler one that you gotta have, uh, those are gonna help you with research or with scraping. That covers a lot of use cases out there, especially some of them are simpler ones.

**[0:09:30] João Moura:** Now, where again, where where value really gets unlocked is where when you can tap into internal data. And a lot of the times, that is kind of like some of that bigger enterprise kind of like companies that we all know of, like Salesforce, um, or other CRMs out there, SAP and a few other things like that, or again, internal grown systems, like those systems that have been running in these companies for many, many years. And they need to tap into this data, either through an API or through a data lake.

**[0:10:01] João Moura:** So I would say those are some of the things that we really look for, um, in terms of tools that really unlock value.

**[0:10:08] Lukas Biewald:** Do you see companies like, for example, Salesforce modifying their APIs or even changing their pricing model due to, um, agents using them as a tool? Like, for example, what does a per seat price mean in the context of a um agent? Do you think if I have thousands of agents and they all access Salesforce, I'm gonna have to buy thousands of agent seats for them?

**[0:10:29] João Moura:** I don't think the seats is the way to go. I mean, we're seeing some companies experimenting with that. I think LinkedIn is doing something, uh, in that in that lines. And I I have heard that, like, it's. It's a tough problem for them to navigate at any other company out there, right? Because do you charge less for an agent seat? And what if this agent's, like, so good that it actually, like, you don't need an actual seat?

**[0:10:51] João Moura:** Like, you're basically kind of blasing your revenue a little bit, so it it might, it it makes for all this kind of like, interesting, kind of like, uh, uh, problems that you got to navigate. Now, I do think there's going to be endpoints that are more focused on agents, and they're gonna become like a like optimize for agents. Uh, that is sad. I think that will happen further down the road than most people, uh, believe. Just because it's so much easier for you to try to get agents to comply with the inputs and outputs that we humans already use.

**[0:11:26] João Moura:** Because then you do that once and you unlock the entire internet instead of trying to do the opposite and you have to translate the entire internet into something that the agents can do. So I think it's gonna take a little longer, but there's definitely gonna be kind of like some, uh, pricing models updates based on kind of like agents now consuming data instead of humans.

**[0:11:46] Lukas Biewald:** And now your website says that 40 of the Fortune 500 uses, um, CrewAI. Can you can you talk about how you got that adoption? And also, um, can you talk about what verticals or industries have the most adoption of of agents in CrewAI?

**[0:12:02] João Moura:** Yeah, uh, I gotta say the adoption is insane. Uh, a bunch of that happens through open source first, and then eventually migrating into enterprise. So it's interesting because when I first created this project, I was not expecting it to blow up the way that it did. Uh, and I'm very thankful for it. I think we have an amazing community now. And if anything, like, um, I still remember being actually in here where I am now. I'm on 15th in San Francisco.

**[0:12:30] João Moura:** And I remember being back in, oh, a room on the other side. And I remember, like, uh, someone from Oracle approaching me, um, back then. CrewAI was not a company yet, it was just an open source project. And I remember this person approaching me and saying, like, Hey, we are Oracle, we're using we're using CrewAI in production. Um, can you help us? We need some help. And that, for me, was a big aha moment because I was like, all right, so this big company is actually using this open source project and they need help.

**[0:13:00] João Moura:** If I want to provide them with the help that they need, I don't think I can do that through an open source kind of like side gig kind of thing. I mean, I did have resources, so might be worth it to turn this into a company. So it was very organic. Um, I think, um, me being so enthusiastic about agents overall and just being a lot of like, doing a lot of being in public also helped to drive some of the adoption.

**[0:13:25] João Moura:** And then a lot of the educational content that follow up that I think also helped penetrate a lot of those the enterprises. And uh, and then nowadays, it's funny, like we have, for example, uh, major banks reach out to me, and I learned that their CTO has been using CrewAI open source. And I'm like, all right, I guess that's happening. But, uh, in terms of verticals, I think there's a there's a again, there's a little bit of favor thing.

**[0:13:53] João Moura:** There's definitely a few verticals that are moving faster. Uh, I think GCP is impressive, how fast they're moving. I think, uh, finance is now also moving a little faster. But again, like they, they move faster in terms of interesting, but it's a highly regulated industry, so it takes like a lot of mark conversations to get them to to start deploying some of these things. But, uh, but I would say those queues are are industries that have been impressing me, and also some insurance companies as well, they're just moving fast and faster.

**[0:14:21] Lukas Biewald:** Interesting, so finance and insurance is actually moving the fastest And GCP, Yeah, and GCP. Interesting. What about, um, in terms of volume of usage? So, you know, you're also your website says over 100 million multi agent crews have run using CrewAI. So, um, what have been the sort of, like, longest running, um, crews, or how? How however you define that?

**[0:14:48] João Moura:** yeah, I gotta say we need to update that number.

**[0:14:51] Lukas Biewald:** All right, well, what's the number now?

**[0:14:53] João Moura:** Well, I can tell you that January alone. I was pulling this because we had a, uh, basically, uh, the presentation yesterday, and I pulled it up then. And January alone was, uh, over 50 million agents. So that was, yeah, that was insane. So I like, lifelong time now, I don't know. I I don't think we got to, uh, I don't think it got to a billion just yet, but we might, we might be closing in. Uh, but what I would say is, um, yeah, it's insane to see, like, some of the scale and and like, how how fast, like, things are are accelerating there.

**[0:15:26] João Moura:** Sorry, I missed. I forgot the question.

**[0:15:29] Lukas Biewald:** Oh, the question was like, what are the longest running agents?

**[0:15:32] João Moura:** Oh, longest running agent. I gotta say a lot of that, like, there is one that we're like co-building with a customer now that will be a long running one. Uh, I think more than the long running ones, the one that I have seen that are more impressive are the crews. They're basically groups of agents that had a lot of agents on it. So I remember seeing at once a like a crew that had 21 agents on it, and that, for me, was was impressive.

**[0:16:03] João Moura:** Like, all right, I didn't see that one coming.

**[0:16:04] Lukas Biewald:** And what? How did? How does that work? Like, why so many agents? What was it doing?

**[0:16:08] João Moura:** Yeah, so this one is great. It was actually, um, it was a company that basically sells reports to bigger companies, like those bigger kind of like, uh, consumer facing companies around their competition and their market positioning. So this company would basically charge them a lot of money, and they would put a huge research on here. All your competitors, here's what they're doing. Here's like photos of their actual stores. Here's like how they're positioning products on their stores in in a bunch of like other like researches, a lot of like market competitive analysis.

**[0:16:49] João Moura:** Now, some of those things you're gonna get people on the streets and and you can replace that just yet. But a lot of that was research was like digging and digging and digging and digging a bunch of materials. And you kind of like, know the sources that you want to check, but you also want to do some exploratory, and that was basically what this use case was doing. So the reason why they had so many agents is that they went very specific and specialized on each agent on what they want, each agent to research.

**[0:17:18] João Moura:** So if one agent would be searching kind of like the financial aspects of the market and competition, there would be one dedicated agent to that. If they had another agent to research something else, like, uh, branding, positioning and marketing, there would be a specific agent to that. Uh, and in the end of the day, they would produce like, uh, tens of pages, lo pages longs, kind of like final reports. Then that's kind of like the the deliverable back to the customer.

**[0:17:46] Lukas Biewald:** What about the different levels of autonomy with the agents? I mean, I know for, like, autonomous vehicles, they talk about the five levels of autonomy. You've talked about human in the loop, like, how how do you think about that? Like, how do you set up agents to actually go back to humans and engage with them? How do you put in checks and balances to make sure that, uh, the agents are producing the intended results?

**[0:18:09] João Moura:** Yes, so I think that, uh, that that's a great question, by the way. I think there's a lot of different ways that you can do it. So depends a lot on kind of like how much precision you want to have and how complex is your use case. For most of the people where they want to have a lot of agency, they want to have like these agents to really go out and figure it out. There's a few different ways that you can put, like check and balances.

**[0:18:29] João Moura:** So that might be number of requests that these agents can make on how how much time they have to get the work done, and those things kind of help them to just kind of like not go haywire, right? Yeah, and then there's a few other things that you can do, like you can add programmatic and like, um, um, what do we call it? Guardrails, so you can call programmatic guardrails. So as an agent, finish something instead of just saying like, I'm done, and you got to get this output through this programmatic guardrail.

**[0:18:58] João Moura:** That it might be actual just Python code if you want to, that might check, for example, I don't know. Let me think about something silly. How many times the word but shows up, right? And if there's a lot of but like showing up throughout the thing, you kind of like, No, this is not good, you go back. So this would be like when, like, a few of the options that you have now. If you're talking about more complex use cases, so, for example, I mentioned the one about filling IRS forms for taxes.

**[0:19:25] João Moura:** That is something that you want to be way more careful about. And by the way, those forms they are in this use case, it was a form that was around 60 pages long, but fear not, it comes with instruction manual, and the instruction has 720 pages long. So, yeah, so, uh, for that use case, again, you you want to have somewhat of agents, but not so much. So what we would do is you would have a flow, so we have a kind, like, I'll create, I flows that allow you to kind of like intertwine agents and regular code together.

**[0:20:00] João Moura:** So you would do basically with regular code, get each page off the form, extract all the questions from that page. Now those questions, one by one, go to a group of agents, and these agents would then perform a rag on an internal database. They would check out a few data lakes, and they would create the answer to answer that specific question. And in the process of doing so, they could look at the instruction manual. So the agents are always working in the scope of one single question, so they're not like going crazy and hallucinating about all those different questions and all that context.

**[0:20:41] João Moura:** And then that is done, and they go to the next question. So this would be one way for you to have way more control, um, on this kind of like more complex use cases. And at any point in time, with CrewAI, at least, you can force human the loop. So you can say, like, Hey, there's gonna be a human in the loop right here, and then the agents can, like, stop and wait for you to get back to them.

**[0:21:01] Lukas Biewald:** I guess that's a good segue into talking about your CrewAI platform. I mean, you've you've given me some really good examples of, um, agents and and why that would be useful. But can you describe, um, you know, what parts of the solution you leave to the ecosystem or other vendors, and what parts of the solution you actually do for your customers?

**[0:21:22] João Moura:** Yeah, I think there's a lot of parts on the vendor sides, like the the one doing anything AI, not even agents, anything AI is very, it's very complex. Like it's not only doing an API call, that only gets you so far, right? And I believe there are problems out there that are big enough that entire companies should exist to solve them. And we're already seeing that, right? So there's like big clusters. I want to say one is, um, other things around fine tuning, like that feels like complex enough that you could have entire companies just to help you get that right, right.

**[0:21:57] João Moura:** Uh, and I mean, you folks have a lot of background of that, right? If you go back, you folks are OG. AI, all the things from hyperparameters like tuning and everything in that, like that. That requires a lot of work for it to do it right. And if you're just doing like Jupyter notebooks on your local computer, you get lost very quickly. Uh, and you did. I still remember not being able to replicate something and being just so, like, so mad at myself.

**[0:22:23] João Moura:** Like, I don't remember what was the hyperparameters that I used. If I can't get the same ratio, but yeah, again, sorry, derailing the conversation, but I think, like, fine tuning is a good example. Like, I think that's something that we don't offer right now. Um, maybe in the future if I ever get there, but they have so much on our plate start now that I think this is this is a big enough problem for a company to solve.

**[0:22:44] João Moura:** Um, so that would be one example of it.

**[0:22:46] Lukas Biewald:** But sorry, what do you do actually? So that's probably how I should have asked the question.

**[0:22:50] João Moura:** Yeah, so there's a few things, uh, on the platform. Again, we try to cover all those different stages of the agent building, deployment, and monitoring, iterating. So you can go into our platform and you're gonna be able to one configure everything for your company. So that will be not only inviting the right people by setting up the right permissions and roles, setting up your LLM connections, and you can bring any LLM. You can use a private proxy if you want to, so you can, like, connect all those things in there.

**[0:23:18] João Moura:** Now, once that you're set up, you go into building, like, all right, I want to build these agents now. So now in the platform, you can not only use the open source, but you can also build with no code. We also offer, like, Crew Studio, where you can basically chat your way into an automation. So you basically got this automation going, the agent's created, and you can deploy right away, or you can go back into code.

**[0:23:40] João Moura:** Uh, so we talked about setting up, planning, and building. Now, once that you build them, you need to deploy it, and you can deploy it right in there as well. That automatically becomes an API, and that's production graded. So I'm talking, uh, load balancer, auto scaling, SSL, like all those different things. Um, so you have this an API now that you can integrate with other things. And then you got a bunch of metrics based on that, and those metrics are more agentic.

**[0:24:06] João Moura:** So you get to see not only prompts and and uh, and and that, but you can also go into the quality of the outputs, the like hallucinations, and how much there's happening in there. And you can also set up custom metrics if you want to kind of, like, track that and set alerts based on that. And then we go towards like the the end, kind of like of the the life cycle, where you want to now iterate on this.

**[0:24:28] João Moura:** And what that means is, oh, a new model popped up, like O3 mini, and I want to test all my agents on that to see what are the ones that perform better. So we offer ways for you to basically, with two clicks, kind of like now you run other agents with a new model. You can see how they perform. You can change that, you can choose and pick and choose the ones that you want to use that versus not.

**[0:24:47] João Moura:** So that makes it easy for you to basically constantly iterate on these agents. So the TLDR would be kind of that. There's a lot that goes into all that, and then there's a bunch of things in metrics, a bunch of things in the iteration. But there's even like, uh, other features, uh, that comes into that as well.

**[0:25:04] Lukas Biewald:** What about memory? I feel like that's like a big topic right now with agents. Do you help with memory at all?

**[0:25:09] João Moura:** yeah, for sure. I think, like, memory, um, when people ask me about it, memory or rag and things like that. Yes, all those they must exist, and you know, agents work better with that. I'm just finding that out that is becoming table stakes very quickly. Like, uh, yeah, you you gotta have it. No, no questions asked. Like, there's many different ways that you can build them. I would say it's easier to build them when you're using crew because you integrate with any other vector database if you want to.

**[0:25:36] João Moura:** You can integrate with any data source that you want to, so that makes it extra easy. Um, but yeah, I think, uh, memory, we do have all the agents in CrewAI, open source in an enterprise. They have short term memory, long term memory, and entity memory. And then there's a fourth type that we added recently that is kind of what I call, um, I think, user memory, and that is basically memory that can preload in your agent.

**[0:26:02] João Moura:** So you can give them like a a preload set of like documents and PDFs and things like that. That goes into their memory from the get go and gets stored in there. But all the short and the long and the entity memory that gets, uh, populated autonomously as they're doing their work.

**[0:26:18] Lukas Biewald:** So how do how does short term memory, long term memory work differently?

**[0:26:21] João Moura:** Oh yeah, so short term memory, as doing the execution, you have like a few agents working together. So this memory is kind of like a sandbox where these agents can share information with each other. They can always delegate specific work to each other if you enable that, but they have kind of like this commonplace where they put like some of their learnings and some of the things that they're doing that these agents that are working together can tap into.

**[0:26:46] João Moura:** But that gets reset in every run. So you do a run and then boom, everything gets cleaned up and a new run happens. Now, the long term memory is where they store learnings that are coming from multiple executions. So because in crew, we force you to say, what is the expected output that you want to have from each task? We basically have something to compare against. So we can actually get the real output that you got and compare with what you expected to get.

**[0:27:15] João Moura:** And then you can see what were the things that the agents did right or wrong, and autonomously create rules for them to follow in the future. Saying like, Hey, you didn't got this thing right, so let me create a like now a validation for that. And and that just makes sure that over many executions, your agents got better and better.

**[0:27:35] Lukas Biewald:** And now do you have ways of monitoring automatically when the agents run for a long time to make sure things haven't gone haywire?

**[0:27:43] João Moura:** Yeah, you cannot only monitor, but you can set specific kind of like, um, hard stops. You can say like, Hey, you can only run for like 60 seconds. And then what happens is as it gets to that 60 seconds, we do a final call saying like, Hey, all right, you're done, your best answer right now. And that either kind of like goes through or kind of like it might blow up and fail, like, all right, you just have to not go over.

**[0:28:04] João Moura:** So there you go.

**[0:28:06] Lukas Biewald:** And now what do you see in terms of, like, open source versus closed source adoption, like, which is which is more popular? Do you have, like, specific recommendations? If I came to you and said, I don't care, I just want the best model, like, where, where would you guide me?

**[0:28:20] João Moura:** Uh, you turn you s you're talking about open source models

**[0:28:23] Lukas Biewald:** Models? Yeah, LLMs, yeah, yeah.

**[0:28:24] João Moura:** like LLMs. Yeah, well, I gotta say, uh, one. I'm a huge believer on open source. I have been living and breathing open source for many years. I had quite a few projects through the years, and I'm a strong believer that in the end of the day, open source is gonna probably win. Uh, whatever winning means here, right? I don't think meaning is like, uh, is like necessarily closed models, right? Not gonna exist anymore. Um, I think it's just they're gonna be more accessible and easier for people to run these things at a scale.

**[0:28:51] João Moura:** Now, the models that are most used are closed source models nowadays, those are the most commons, both in our open source in our or, um, our enterprise, uh, product. But we are seeing things take up on the open source. There's a lot of people that are running agents locally using things like Ollama. Um, it's it's very usual for us to see people doing that as well, and it's more common in enterprises when you're talking about highly regulated industries.

**[0:29:21] João Moura:** So, for example, finance and insurances, they usually want to self-host their models because they wanna run the whole thing in kind of like a airtight container, right? So no information is going out just because they need to be very mindful about their data.

**[0:29:37] Lukas Biewald:** Have you started to see R1 taking off in your metrics?

**[0:29:40] João Moura:** I gotta say I'm getting a lot of mixed signals. Uh, yes, there are people using it, even in highly regulated industries, what I think would be a hard like no, no, but, uh, I think, like, uh, it's interesting to see, like, some, uh, some of these companies actually starting to play with that and see how that goes. Now, if that are really gonna take off in in the US specifically, I'm not sure. Uh, I'm not sure.

**[0:30:02] João Moura:** I think it's gonna depend. Now, what I'm grateful for R1 is just the idea that people now it's almost like they believe that open source can go even further than what they I think they mostly thought. So that's already inspired a bunch of people on trying new things, and I guess in no time, we're gonna start to see a new influx of models, especially cognitive models, coming up from open source initiatives. So I'm very curious to see how that's gonna play out.

**[0:30:27] João Moura:** But yeah, even the distillations that we're seeing now with Crossi crossbreeding, R1 with other models, I'm finding them to be very interesting.

**[0:30:35] Lukas Biewald:** How would you describe the boundary today of what agents can do and what they can't do?

**[0:30:40] João Moura:** Hm, that's a good one, I think. In the end of the day, there's I think there's more that agents can do than most people would think of. But it's not enough for people to be worried about, like, uh, losing their jobs or anything like that. I don't think we're or I don't think we're close to that just yet. Um, but there's a lot that these agents can do. I think where if you take one step back, like the value that companies or people in companies get from these agents, it's basically a math formula in my mind.

**[0:31:13] João Moura:** Whereas how complex is what these agents are trying to do times how much autonomy they have on doing it? So if you can automate your most important process in your company and they can do it completely autonomously, that is a lot of value, right? And that's why you see a lot of people aiming for kind of like co-development, because that's a very crucial process in the company. And if these agents can do it autonomously, that adds a lot of value.

**[0:31:44] João Moura:** I think we're not on that top quartile just yet of like those most amazing processes with like no hand holding, and I don't think people are ready to have no hand holding. Um, in like the better half of the processes so far. Now, what we are seeing is agents that work with humans, and that seems to be taking more crash, like more, um, that seems to be like basically speeding up and getting more adoption. Where we basically see where before you had four people doing something, you now have one person that is overseeing agents, and the other three people got reallocated to do more interesting stuff because we're doing kind of like busy work.

**[0:32:27] João Moura:** So I think right now it's gonna be a lot of efficiency gains, so repetitive tasks that it went out to me. I think we're gonna see an influx of these agents going into decision making, probably in the next, like, uh, couple years. And that's where I think we're not going just yet, and I think it's gonna get very interesting once that we do.

**[0:32:47] Lukas Biewald:** But now, like, what problems are hard for agents, right? Because, you know, like, if you, you know, if you just were dropped down on planet Earth right now, you'd be looking at this like, Wow, you know, these agents can do, um, you know, kind of Putnam level math problems. They can, like, pass the LSAT. Like, you know, if I had, like, an employee that could do that, I'd be like, Wow, you know, you could probably, you know, do a lot of the things around here, like, what?

**[0:33:10] Lukas Biewald:** W- Where do you see them break down?

**[0:33:13] João Moura:** Yeah, I think, like, there again, what we're seeing these models like really excel, especially cognition models, is on the math problems is because they have a hard, right answer, right? So they can do reinforcement learning based on that, and that's why you're seeing a lot of those models succeed on that. But the real world, especially in the business, is way more nuanced. So, um, yes, in the end of the day, you can say, well, that was a right choice.

**[0:33:36] João Moura:** But at the point in time where you're making a decision, is way more nuanced. Where, like, there's an element of kind of like taking a bad or kind of like d deciding what you believe is gonna be like the best. So I think, like these agents, like, they're getting very good due to some of the techniques of reinforcement learning to answer problems that have tough, hard, yes or no, or specific number kind of like, uh, questions.

**[0:34:01] João Moura:** I think they're not as well on, uh, as of version nuance just yet. I think you can use some of that, but I think it's harder for people to believe it, uh, that this was exactly the right choice, right? So I think that's kind of like where we're seeing the gap now. Even though if you had the human that was amazing at kind of like doing those math problems, you would assume that that would translate into kind of like a logical capabilities that would lead to better decision making on nuanced situations.

**[0:34:30] João Moura:** But that's not how this model works. They, uh, they work better in that situation because they have the reinforced learning that optimized them for that. But that doesn't necessarily means that they are better on these nuanced situations just yet.

**[0:34:41] Lukas Biewald:** And now do you see that changing with, like, um, o1 and o3, like, do those feel like big, um, improvements in that regard, or or not yet?

**[0:34:52] João Moura:** I think not yet. I think they're amazing models, right? They're amazing models, definitely very capable, but especially for agents, we're not seeing they take off that much. In the end of the day, I think, like the same, um, engineer's first principles apply here, yeah, and that is what is the minimum that I can get away with, right? And, uh, that kind of helps you to make sure that you're optimizing for speed and cost and all those things.

**[0:35:15] João Moura:** So I think a lot of people are still running agents on, like, in case of open AI, like for O mini, for example. And if that got you what you need, then you might be better off than a bigger model that will be more expensive and a little a little more slow.

**[0:35:31] Lukas Biewald:** Now, um, I want to talk a little bit about your your architecture. I think you sort of visibly took, um, lang chain, kind of out of your out of your stack. Can you talk about why you why you did that?

**[0:35:43] João Moura:** Yeah, sure, I mean, there's a few different things, right? I think, uh, on the early days when we started CrewAI, I think using LangChain was a good, a good way for us to get access to a bunch of tools from the get go. And because those tools would help a lot of those people, kind of like, take some of those actions, I think that was kind of like, all right, let's let's do LangChain here because that will help us with some of that.

**[0:36:05] João Moura:** Now, as we start to grow and we start to build more and more logic in these agents, it started to get very hard to keep things in sync. And some of the decisions were starting to diverge and how, like we were taking some of the decisions on the framework, and how, like, how they were like changing some of the code on their side. So it got to a point in time where we were overriding like 90 of the code that we were importing, and that was one specific piece, the class that we were importing.

**[0:36:34] João Moura:** But we were overriding so much of it that every time that we need to update a version, we need to, like, go to crazy rebases. So I was like, all right, this, this is not gonna fly, like, we need, we need to cut this off, this is not working. And I think it was the best decision that we ever made because it helped us with basically growing to other things that were constrained before because we had to overwrite certain methods in certain way, but now we can do way more.

**[0:37:00] João Moura:** So, uh, yeah, that's like, that's one of the more technical reasons why we decide to go that route as well. But there are also commercial reasons that helped us kind of like make that decision. And some customers that were having issues with that dependency specifically, so all that kind of like end up contributing to the decision.

**[0:37:16] Lukas Biewald:** Are there integrations that are important to you that, you know, you, you have that you're definitely gonna keep, or, um, integrations that you want to add over time?

**[0:37:24] João Moura:** Yeah, I think the overall tools ecosystem is something that, like, the more cons, the more like, um, the more you connect to all these other things, the better. So, for example, you still can use any LangChain tool with CrewAI, but now we don't depend on that. You can also use any LLM index tool with CrewAI, and we don't depend on that. And then you also have the CrewAI tools that you can use itself, and you have amazing companies out there like Composeo that has a lot of tools.

**[0:37:52] João Moura:** I think they have like over 300 tools that you can also use with CrewAI. So I think we're finding a lot of value that these kind of like dependencies really help, so we don't need to rebuild things over and over, kind of like you build it once. Um, so I think that's something that we're gonna not gonna like change anytime soon. If anything, I want to go deeper on some of those integrations.

**[0:38:13] Lukas Biewald:** What, um, what AI tools do you personally use day to day?

**[0:38:19] João Moura:** I'm, uh, I'm a I don't know if it's a hot take, but a lot of people are like, on the wind surf versus cursor, kind of like, uh, kind of moment. I'm a cursor guy. I love cursor.

**[0:38:30] Lukas Biewald:** Okay, nice, all right, me too.

**[0:38:33] João Moura:** It works so well for me. I try wind surf, and I think, I think, honestly, this is not a dig on windsurf. I think they're doing amazing work. But the thing is, uh, I was doing, I'm using cursor so much that I just got the hang out of it. I know how that thing ticks, and, uh, I think like moving to something else was just like, uh, just make my life a little harder. Like, all right, I'm not ready for that.

**[0:38:58] João Moura:** But, uh, a lot of chatGPT, um, and tropic, like the the Sonnet models have been in and out, at least on the UI. Uh, it has been working on the I have been using a lot of the coding side of things, but other than the coding, the actual models interface. I don't, I don't use it that much anymore, but a lot of ChatGPT, a lot of cursor, a lot of crewAI. We do

**[0:39:22] Lukas Biewald:** So you use cursor with Sonnet

**[0:39:24] João Moura:** we use cursor with Sonnet. I'm using some O3 mini now as well, and I do find a few use cases where I like it. But Sonnet's just like, I don't know what black magic they're doing over there, but that thing works.

**[0:39:35] Lukas Biewald:** with Sonnet? Do you have any, like, um, is there any features of, uh, like, lesser known features of Cursor that has made your life easier lately?

**[0:39:46] João Moura:** Uh, I think, honestly, I use a lot of, like, what people use. Uh, the one thing that I start using that I think people most people don't know is that you can add custom cursor rules, and I think people don't realize that, and you can do that per project. So you can create a file that I think it's called, uh, doc, cursor rules, and you can throw strings in there to dictate custom rules to preload into cursor given the project.

**[0:40:12] João Moura:** So, for example, if you're doing a Python project, you can add a bunch of, like, Python rules, uh, in there. They're gonna be embedded into your prompts. So that has been some things that I didn't know about and that I start using it now, like everyone on the teams does, and I really like it. I didn't know about it.

**[0:40:29] Lukas Biewald:** Cool, um, any other tools that you that you love right now?

**[0:40:33] João Moura:** Uh, let me think about tools that I love on the AI. I gotta say I tried Perplexity for a while. Uh, was not a huge fan. Um, I'm now starting to use more of the deep search stuff. Uh, I'm finding that to be pretty good. I think OpenAI did a good work in there, but I think I think that would be the book of it. I use it some stuff. Oh, no, I use some stuff for video editing as well.

**[0:40:59] João Moura:** There's a tool called Descript Amazing video editor. I mean, honestly, the best video editor ever. I think this might be a hot take for people that are really into video editing. But, uh, for me that like, I'm just trying to produce content. I want to make sure that everything looks good. It's easy to cut, easy to caption and everything. The script is just like chef's kiss.

**[0:41:22] Lukas Biewald:** Nice, all right. Um, how do you, uh, how do you respond or think about concerns about AI replacing human jobs?

**[0:41:32] João Moura:** I gotta say, I think in the end of the day, um, what I tell people when people ask me about it is that we don't like to assume that, but we have way less control over our futures than we like to believe in. Like, honestly, w- you don't know what's gonna happen tomorrow, like no one knows, everyone's kind of like going through this for the first time. So what I would say is you got a couple options.

**[0:41:55] João Moura:** Once that you got to that, either you can choose to focus on the things there are in your control, or you can be sorry about the things there are not. So if that's not something that is you on your control, and it's it's just not, well, then what is? And in my mind, the choice that people that are worried about that are gonna make is like, well, maybe I should learn how to use these tools. And again, that might not be enough, or that might.

**[0:42:27] João Moura:** But at the end of the day, I would rather be someone that knows how to use them than not, just because I think that statistically, that would put you in a better position. So what I what I try to reframe the conversation in there is like, hey, maybe this is the way that you should be thinking about it.

**[0:42:44] Lukas Biewald:** Do you think that technical skills will help with building agents, or do you think over time, agent building will completely move to a no code type of mode?

**[0:42:55] João Moura:** I think long term, no code's gonna become more and more popular and probably get a lot of the market share. Uh, that would be my my take now, just because it it's not that people are gonna get like, lazy or anything, it's just that there's way more people that don't know how to code than people that do. And eventually, if you want this to be running on, like every company out there, you're gonna need to support that super well.

**[0:43:17] João Moura:** Now, I do think that coding will still allow more customability, and that's gonna gonna that's what's gonna lock most of kind of like the high impact use cases in the end of the day. Similar to coding, if you think about it, like, can you build a website with no code? Yes, you absolutely can. Is there amazing templates that are gonna make like that good, super good? Yes, there are, but like, if you wanna do something that is super custom, do you need an engineer?

**[0:43:43] João Moura:** Yeah, you do.

**[0:43:44] Lukas Biewald:** What do you think about your product? Do you think it'll ultimately be aimed more at a no code audience?

**[0:43:51] João Moura:** I mean, we do have no code tools now that allow people to build in. I think they're gonna keep investing on that for sure. Um, just because I think there's a there's a broader audience that we're trying to serve that goes beyond regular engineers. And not only that, I'm finding myself using a lot of the CrewAI studio just because that makes me faster, similar to kind of like cursor and whatnot. So, um, I want to say that, yeah, we're probably gonna keep investing on no code as well.

**[0:44:20] Lukas Biewald:** Hm, interesting. Well, what applications have worked well for you, um, inside your own company?

**[0:44:26] João Moura:** Well, in my own company, the thing that we use the most is Crew Studio. So, um, Crew Studio is basically, uh, we took a very different approach than most no code out there. Uh, I have, uh, an opinionated view that I don't love node UIs. I know that's kind of like what a lot of people can like use for kind of like a no code. I find that it makes for beautiful screenshots, but if you're dropped out of nowhere in a page with 30 nodes, it's impossible for someone that doesn't have, like, a technical knowledge of it and understanding what is going on, uh, with ease.

**[0:45:03] João Moura:** Now, what we decide to do to kind of like offset that, and you can go into app dot create dot ai dot com and create a free account and test it out, is we have a what I call a gradually evolving UI. So you start with a chat interface, similar to, uh, custom GPTs back in the day. So you're start to chatting your way into what you're trying to build, and that is kind of like building that mental model in your mind.

**[0:45:34] João Moura:** Once that you're ready, there's a button that pop off that you can click that's kind of like, all right, generate me a crew plan. And once that you click on that, the UI evolves into a tabular view where you can see your agents, you can see your tasks, you can still change them, but now it's more structured. It's not like plain text, and you can still keep chatting if you want to. But once that you get there, you got a new button to generate the crew, and that takes you to a node view.

**[0:46:04] João Moura:** So by the time that you get to the node view, you kind of like you have a full understanding of kind of like how you got there and what you're trying to build. It's easier for you to customize it that that way, so that's kind of like our approach to it. And then from there, you can download the code if you want to.

**[0:46:18] Lukas Biewald:** But for you personally, inside your company, what's working? Like, what are you actually using crews to to accomplish?

**[0:46:24] João Moura:** Oh, sure, uh, so we're using a lot of things. We're doing generating marketing content automated. We are doing a lot of meeting prep. So agents that are automatically researching people, companies, everything along the way. Uh, anything that is related to PR reviews, we have crews reviewing every single pull request across the company, open source and closed source. That has been pretty good. Uh, agents doing a bunch of support, so that is also working pretty well, writing responses and all that.

**[0:46:53] João Moura:** Agents that are picking off, uh, basically recording and transcripts of meetings and following up with next actions, presentations and such. That is also pretty good. Uh, and one that I love is one that happens with onboarding. So as these customers onboard the platform and they put their names and emails, we kick off agents that research them, but I don't stop there. Based on the research, these agents infer what they might use agents for, and then they get those ideas and push them to two different places.

**[0:47:25] João Moura:** One is our CRM. So now every marketing like engagement that they get is super custom on what we believe that they might want to build. And we also inject that into the product itself, so the product self-customize based on our agentic hypothesis of what agents they will build. Uh, and then it just works, and they're left wondering, like, how this app knows that I'm trying to go that way. And that's kind of like some of the how the sausage is made.

**[0:47:58] Lukas Biewald:** Interesting, do you have any, um, predictions for 2025? Any any new applications that you think will become possible with better LLMs or better tools?

**[0:48:07] João Moura:** Uh, I think I I think we're gonna see some of fine tuning coming back. I think like people have not been talking much about it. I think we're gonna start to see some fine tuning coming back, especially because smaller models, I think they're gonna start to trend again. Like, some of the things coming out of the R1 is like, Hey, can we reuse this to make, like, small models amazing? And what would that look like? Uh, I think that's gonna be a trend this year where we're gonna have smaller models that are very, very capable, like better than like 70 billion parameters models nowadays.

**[0:48:38] João Moura:** And they're gonna be able to fine-tune those with ease to use them to run agents.

**[0:48:45] Lukas Biewald:** All right, well, Joe, thank you so much. This is a real pleasure to talk to you.

**[0:48:48] João Moura:** Thank you so much for having me. It was a lot of fun. I really appreciated Lucas, and, uh, this was great.

**[0:48:53] Lukas Biewald:** Awesome. Thanks so much for listening to this episode of Gradient Dissent. Please stay tuned for future episodes.