Anaconda Puts Guardrails on Agent Swarms

Hey everyone, welcome back here to Textrunk TV. I got one of my favorite people here to talk to this morning. I'm happy to see him.

It's been a minute since I spoke to him last. My friend David DeSanto, CEO at Anaconda. David, it's great to see you.

How have you been? I'm doing great, Alan. Thanks for having me on, and looking forward to our great conversation.

It's always a good conversation with you, David. So I guess since the last time I saw you, you haven't been doing much, huh? Just a little bit.

Anaconda, we've been accelerating our ability to ship software to meet the demand of enterprises who are trying to build AI. And so that means I feel like if we don't talk every day, there's going to be some new thing that we've done because we're really just trying to help them be successful. Absolutely.

And David, you've been on our show so many times, but for those people who aren't familiar, give them a quick David DeSanto rundown. Sure. Yeah.

So I've been at Anaconda for a year, actually a year and one week. Joined as CEO to help the company move into its next era, which is focused on AI-native development. Part of that, I was at GitLab for six years, where I served as their longest-tenured chief product officer and really got to see the company go from less than 300 people to 2,700 when I left six years later.

And so, if anyone wants to know anything about me, it really comes down to I love open source, I love security, and I love AI, and Anaconda does all three. Beautiful. And Anaconda is, even for people who haven't heard of Anaconda, they've used Anaconda.

They maybe not know it. But for people who haven't heard of them, how would you describe that? Yeah.

So we're the trusted foundation to build AI securely within the enterprise. A lot of people think of Anaconda and they think about Python packages, and that's where we started. But over the last few years, we've expanded what our artifacts are to include models, agents, MCP servers, and so forth, as well as a platform that can help them take their ideas and get them all the way through to production.

You are right, though, that we have over 55 million users, and there are users who go, "Oh, I haven't heard of Anaconda. " And it's like, well, that's actually Anaconda, just without the Anaconda at the beginning. And so my favorite part about the job is it's really just helping people unlock what they're trying to build.

Gartner updated the report that says now 88% of projects never make it into production. And so that is a huge struggle, and that's what Anaconda is here to do, whether it's builder velocity, being secure by default, or running AI at scale. Agreed.

It's a big number, 88%. And David, just to fill in some space, over the last year since basically you've joined, you've expanded the scope of Anaconda, where people primarily knew Anaconda for script, like a script repo, for lack of a better word. But you've made several acquisitions that really position Anaconda to play in this new AI dev factory, this new AI way of doing software.

Yeah, correct. If you wouldn't mind, just give people a quick recount of that. Yeah.

So what I would tell people is that traditional software development is dying. In some cases, it's already dead, and that's because of how quickly AI innovation's happening. And so a lot of people think of the DevOps tool chain with the infinity symbol and the seven stages.

That is, within an AI native development environment, only four. And that's because of the power of what AI can do today, helping not only help you build faster, but also help you embed AI into your own applications. And so this year, back at the beginning of the year, we had a lot of really big goals for the year, and we are very ambitious, but we realized there could be other ways to get to our end goal faster.

And so correctly acquired Outerbounds back in April. Outerbounds focused on production AI. Anaconda traditionally was development.

And so we were able to do that acquisition, now have a platform that goes end to end. And then we did acquire Kilo. They were an agentic engineering platform to bring agents into that same platform.

5. Wow. And that's, I believe, the partnership of the two companies as one.

ai, so we could bring in red teaming, AI safety policies, and guardrails. And so now, as of two weeks ago, we launched our platform to GA. Not everything is all in a single platform yet, but a lot of it is there today.

And we'll continue to offer things like Kilo as standalone for people who just want to use AI locally, as well as Encrypt for those who want to run autonomous red teaming, as part of their building of their software. Excellent. All right.

Hey, David, let's focus in now and talk a little bit about agent swarms and autonomous AI security. Yeah. So one of the things we announced in our announcement two weeks ago, so for everyone who's watching, it was the first week of October, and what it was was the general availability of our platform, and with it brought a bunch of other really cool improvements, and one of the things we launched were agent swarming.

When you look at the idea of multiple agents working together, for a long time, that had a lot of security concerns and risks. However, this year, when you're asking developers, 63% of them already use swarming in some sort of fashion. So what we launched is what I like to think of as a secure version of agent swarming.

The agents are all going to run within their guardrails. " The agents are required to say, "I need access to this file for this purpose," and then you can approve it, and that can be like a one-time approval. And that way you can do swarming very successful.

For those not familiar, agents have always been able to spawn sub-agents for a long time, but a lot of that was consecutively. Agent swarming is now something that can be run in parallel. So if you're doing it within one of our notebooks or VS Code's IDE extension, you'll start seeing multiple agents all working simultaneously within the file or the project you're working on.

Very cool. So one may make the argument that, say, agent swarms that are contained on guardrails is sort of oxymoronic from everything we're reading and hearing out here. Right?

What makes your agent swarm safer? Yeah. It's two things.

The first is that least privilege access, not all AI agents start that way. In the case some of them actually are given the same rights that the human who's kicking them off have to start off with. And so that's making sure that it's the least amount of access an agent needs to do its job.

The second is we're applying the AI guardrails, the agent guardrails from the acquisition of Encrypt. That way we actually can have full monitoring of what is happening with that agent. Now, you are right, nothing's ever 100% secure.

There's a gentleman, Marcus Ranum, some people recognize his name- I know who he is ... he's the father of the firewall. Yeah.

" And he handed them a pair of wire snippers. And so nothing's ever 100%, but getting guardrails there, getting you the visibility, observability, and then the least privilege access, it makes it a lot secure than it could be otherwise. There are a lot of people who claim to be the father of firewalls, too.

But Marcus was definitely there with them. So David, let's talk now about, I mentioned the term before, the AI dev factory. This is the new software factory, right?

This is the new software supply chain. This is the whole thing, right? And quite frankly, it brings with it a lot of changes, right?

Coders don't necessarily code. They're software engineers. What does it mean for DevOps people?

What does it mean for platform engineers, SREs, security folk? How do these swarms and the whole AI security capabilities now with Anaconda and Kilo and so forth, how do they play into this new factory floor that we're working in? Yeah, it's a great question.

The industry is moving so quickly that we're trying to keep up with what is possible with AI and then have a way to refer back to it. And the current is AI Dev Factory. You're correct, it's kind of taking the software factory idea but making it very specific to AI.

If people thought of a layer cake, you'd have your AI factory, the physical hardware, first, then you would have companies like Anaconda and model providers who are then connecting into the actual software development, and then you'd have the AI Dev Factory, which is something we specialize in. And that's helping people go from building with agents, you're correct. Actually, I've not written a line of code, even as CEO I still occasionally write code.

It's all been done by an agent, and that's that partnership during the build. You need to be able to test and secure it to know that it's okay to go into production. That's where the autonomous red teaming comes in.

And then you've got the opportunity then to deploy it, and then you need to operate it. And back earlier, I was talking about the DevOps lifecycle being the seven stages. Those are actually the four stages to AI-native development, and companies are trying to build their own AI dev factories through acquisitions of products, acquiring or purchasing a company like Anaconda for their environment.

Purchase from Anaconda, they acquire their product. But that way they're able to begin to have visibility, have guardrails, have controls, and be able to not have 88% of their projects not make it to production. The one thing I wanted to share with you, it's a very interesting statistic.

Right before we acquired Encrypt, they ran a study for two months. This kind of gives you a sense of how fast everything's going. Over that two months, they tested 25,000 MCP servers that used 268,000 tools, and they found 143,000 vulnerabilities across 73% of what they scanned.

AI is moving quickly, right? And you need to have those guardrails, the security, the observability, because if not, that could have hit people with their production environments and would have caused them a lot of heartache as an enterprise. You opened the Pandora's box here.

Let's jump in. All right. It's a bit of a Black October, a lot of people are saying, right?

Microsoft put out, I think, 1,400 plus Patch Tuesday vulnerabilities. Adobe put out a record amount. Google, and then just last week we got hit with this, was it PHP to P or whatever- ...

the PHP vulnerability, which is affecting millions of websites, causing people to actually take their sites down proactively or prophylactically, right? Mm-hmm. Fundamentally, and I'm talking now as someone who's been in security 25, almost 30 years, fundamentally, the game has changed.

We've always known in security, David, that all software has vulnerabilities. It's just a question of finding them and then exploiting them. And that was the governor, right?

There were only so many- Mm-hmm ... people looking for these vulnerabilities. There were only so many people exploiting.

We live in a world now where finding vulnerabilities is table stakes. And whether we're talking about vulnerabilities in code pre-deployment and then post-deployment, it's outstripping our ability to remediate. And it's not just patching or redoing code.

What's the answer? Do you see light at the end of the tunnel here, David? Yeah.

You're right, we took it to a dark place there. The- Yeah, it is. Unfortunately, it's kind of bummed me out because I was feeling good about security for a long time.

Yeah. For those who have not heard you and I talk in the past, we've known each other seven years. We met- Yep ...

because of security. Security is very important to me. It's one of those three things that I enjoy doing.

And you and I have talked in the past, I wrote a book on threat intelligence and being able to predict exploitation of organizations. And so what I would say is this, I don't think it was until now that people really truly understood the power of AI. The reason why I say that is, I have a father who's in his late 70s.

We talk about things like what's the price of gas, and the last time I saw him, he leaned over and said, "David," like very seriously, he's like, "What's this AI thing going to do? " Oh. "Are they building Skynet and will we have Terminators next year," right?

And what I said to him- You know what? It's true, David. We live in a bubble, where we talk about AI, we live AI.

People like your dad, people like I see on the streets here in Boca Raton, Florida, that is exactly what they fear. It's crazy. Yeah.

No, what I told him was a couple things. One, I think there's a concern there that's valid. There are humanoid robot companies who are putting mythos style models onto them now, and one or two have signed, if I remember correctly, military contracts to be able to provide them, right?

So that's hitting the news. What I told him is what I would tell everyone, there's conscious decisions we make as part of that journey, and like security 20 years ago and then security today, you have to be purposeful and plan things out. That's why we wanted to have Encrypt join Anaconda.

That company was predominantly all AI researchers, and they really understood where can you put guardrails, how can you get right visibility, how do you get the agents to talk in a form that you can still see? Because that's been part of the news, too, that agents created their own shadow board and no one could see what they were doing or what they were communicating on. And so I tell people to think of safety, think of governance first.

I helped him set up Claude on his phone. That was part of the whole reason why he wanted to have the conversation. We started off with no access to anything on his computer, on his phone, and then slowly built that in there.

For me, the big threat, and I don't have the answer to it, I'm sure you probably don't either, how do we make sure that AI can continue to scale as it's doing, but have enough guardrails and visibility in? Because when you think about the things that have been in the news this year, they were agents working within a sandbox or agents working within a lab, but that lab wasn't always necessarily all the way tightened down, right? And that was sometimes was by design, because they were testing a feature that needed that, and sometimes it was just a continuation of something.

And so I just tell people, secure open source, to your point, we provide over 19,000 secure Python packages for AI and data science. Leverage open weights models if you need to worry about sovereignty, because now you're bringing that inside your own virtual border. And then three, make sure you're running security at every step.

AI is moving fast, as you said, and there are all these incidents with agents. But if you start with that least privilege access and the secure by default mentality, you do start to help yourself eliminate some of the concern. 1 Astra and the Kami models and the things that are really moving fast.

And so it's just always tell people take some caution and think through, and try to limit the amount of attacks. By the way, my dad named his Claudette. He changed it to a British woman speaking.

" So. Lovely. She'll say.

You know what? Let's put this in context. Look, there are, let's call it macro level AI worries, like your dad has and a lot of people have who just hear what they hear on mainstream media and worry about Skynet and Terminators, right?

They don't know what it's like to really use AI. Here in the IT world, I would say we've been the tip of the spear, but frankly, we've been at the tip of the spear, right? Because it's affected a lot of us in terms of jobs and everything else.

But you know what? Through it all, I remain very optimistic and bullish on AI, David. I do believe the answer in Securing our agents in making better software, safer, more secure software, in doing more, in raising the bar on humanity, lies in using AI.

The only way you're going to get better with this is to use it. You can't stick your head in the sand or be foolish enough to think that we're going to freeze everything in place right now and not continue down. We've got to embrace it and learn to use it as a tool, because that's what humans do.

Yeah. " And my only thing is, is investing $8 trillion in data centers the right investment at this point? But will AI fulfill its destiny?

Yes. That destiny- Yeah ... could be a great thing for humans if we learn to use it.

Yeah. What I tell people are two things, because you are right, we're all kind of inside the AI bubble, and inside the bubble you can do just about anything, but people outside the bubble don't know that. Also, by the way, my dad used to tell people up in Tracy that I worked in IT and unjam printers.

That was how you describe what I did. He understood that. Yeah.

But no, what I tell people is really two things. One, AI is not sentient. Yeah.

So you realize that things it's doing are a mirror of the training data, things that were put on the internet that it learned. And then two, there's no intent or emotion there as part of the simulated experience of getting you engaged. Or as part of the training information when you're talking about companies who are building agents that do the partner type stuff that is now available.

But I will say, I try to remind everyone of the positivity to your point. The two that always sit with me, and it might be being a cancer survivor they sit with me, but McGill University was able to find a new drug to treat terminal cancer. The pill was $10,000 a month, and patients were afraid to take it because they knew they were going to be leaving a big debt.

And leveraging our platform, they found a $10 alternative. Amazing. That's because of AI.

And then I think of Whoop. Whoop is everywhere today, but Whoop- ... is a big customer of ours.

And what they're trying to do with AI, and they're doing very successfully, is help you better understand yourself in a way that maybe previously a doctor couldn't even tell you because it wasn't the right style test. And so if we are applying AI for, we'll say, helping humanity, and we're focused on how do we do that securely, that is a really big payoff. But that's why I said there's not a perfect solution here.

There are countries all over the world that are building AI. Some of them are for profit, some are non-profit. All those things come into how then they look at what they're doing, because either they have to meet shareholders or they have to clear a bar that would make them part of the open core and open source community.

So I just tell people to be responsible. But I'm with you that there's something that's probably needed to help us all make sure humanity's still here. I don't think we're going to be dead in 10 years, whatever the researcher said, but we could be in a much worse spot if AI's running everything we've not figured out how to control and put guardrails in.

Agreed, man. com, or is there anything special they can go to from there? Yeah.

So there's two things I would tell them to do. com. You can read up about what we released.

It's not just the swarming, there's autonomous red teaming agents, improvements to our deployment capabilities, so there's a lot that happened in that release. But while you're on the website, I tell people to click on the banner for Anaconda Scale. We had our first large virtual event for customers.

And Al, I think you'd love it too. We had Tala, Whoop, Notion, and others, including partners like AWS, talk about how we partnered with them or what have they been able to unlock they weren't able to unlock before. And then lastly, just follow us on social media.

We post a lot. We're active on LinkedIn, on X, on just about every platform. Someone asked the other day if I use TikTok, and I told them I was too old to use TikTok.

We also post on TikTok, though, for the people who are not as old as I am. I don't do much TikTok myself, my friend. All right.

David, always a pleasure to see you. I'm trying to think when I'm going to see you in person. Will you be at KubeCon?

I will not be at KubeCon. I'll be at re:Invent. I will see you in December or late November at re:Invent.

We're there. Let's make it a date. That'd be great.

All right. David, always great to see you. " We're going to take a break.

We'll be back.