Unaligned #15: AI Agent primer with founders of Agent Ops AI, @AlexReibman and @AtomSilverman

Yeah.

So if you've ever been in the debugging doom loop where you've been adding print statements to your code till the wee hours of the morning, or you're just not understanding what your your program is doing, which is commonplace for all developers, this is definitely a tool you should be looking at or considering.

I'm happy to announce that I am now taking sponsors for unaligned. I'm proud to have AI Top Tools as our first official sponsor. One of the most important things for me running my podcast is staying on top of the industry, knowing all the players and what they can do. I made a complete list of companies in the AI industry on x.com but no one has as comprehensive a library of tools as AI Top Tools. They break down products by use case like productivity, media, chatbots, customer service and more. A lot of companies are looking for the exact right tools for their needs and there's no better place to find it than AI taught tools. Today we're going to have a little bit of discussion about agents. AI agents are hot right now and I'm real pleased to have you two on. Who are you?

Yeah, I'll start.

So my name is Alex Reedman. I am traditionally a data scientist slash machine learning engineer. Spent a handful of years working in the cyber security industry, basically building machine learning based anti malware. So this is the era of kind of the first few waves of ransomware when that was a thing kind of transitioned more into mlops and building production scale data pipelines for large banks and, and financial institutions. So somewhere along the way I met my fellow co founder Adam. We've just been bouncing ideas for years and we realized kind of with the advent of ChatGPT and how sophisticated large language models were, you could do a lot more than chat with your PDF. And that's when we started building our own agents. And I'll let Adam take the rest of the story.

Yeah, awesome. So I come from like a business background but always saw like in my job that I was doing that there was a lot of aspects of the job that were really repetitive and tedious and, and these were aspects I thought could be replaced. And early in my career I tried to find automations but to say the least there wasn't necessarily very easy ways for non technical people to actually access these automations. In March of last year we actually built our first agent out. It was a data entry automation agent. So it would look at a PDF, would parse information from it and it would upload it to an Excel file or an Online form and this agent, it worked really well, like 90% of the time, but in 10% of use cases we were getting these nasty support messages saying that it was broken on like a dropdown form or broken on certain aspects on the web. And that was actually the inception of agent ops. So we ended up building a debugging tool in house to actually fix this agent, to try and make it as reliable as possible. And we quickly realized that the actual market needed a tool like this and that like observability was very important. Debugging was super important. And the compliance surrounding AI agents is actually something that a lot of companies are missing when they're trying to bring their agents to enterprise. Yeah, yeah.

One thing to add is that it's not just about observability of agents, which is a huge question and issue issue. It really boiled down to the explainability. So our agent was breaking in a lot of places and we, we had no way of understanding why. Sometimes it was core LLM logic, sometimes it was actually errors with the script itself, but more often than not we just had no kind of understanding behind the reasoning behind how this agent solved his own problems. And that was really how we arrived at this problem.

Yeah, let's back up to 40,000ft and talk to people who don't yet use agents. Why do they need to use agents? And what's the state of agents right now? And talk to people. Why? Why do we need agents?

Okay, I can tell you this. I think from our perspective, there's a lot of tedious tasks they do as part of your day to day work that are not necessarily human tasks. And going back to the first agent we built, like data entry, automation, like data entry is not necessarily the most human task. It's not something that like somebody wakes up in the morning and wants to do for eight hours a day. This is something that's a requirement of their job and something that a robot's actually significantly better than a human at doing. So there's certain workflows that people do every day think about anytime you are really interacting with your keyboard, if you're pressing, you know, command C or control C, command V or control V, like there's, there's a lot of like copy paste hell that people have to do as part of their jobs. And this is just like one example, there's a lot of workflows that robots are actually a lot better at completing than humans. And these are the aspects that I think a lot of the agent developers in the space are really honing in on, are there aspects that can actually make people more human? Are there aspects of the work that are very robotic in nature and don't actually need human intervention? I think long term, if you can, if you're at a company and you're trying to innovate and try to think about what parts of my job are the most repeatable processes, these are the types of things that, you know, there's huge innovation opportunities on because it actually, like long term, makes humans more human. There's like the tasks that people don't really need to do and if you can take them off people's daily plate, there's a huge opportunity there.

Like, like go and review two companies. You know why you might have an agent do something like that. What, Tell us what, what people really like to use agents for. What, what are the top three or four things that people are broadly using agents for?

Yeah, I can name a couple. So number one that we've seen a lot of is actually doing research. So just as you mentioned. Yeah, so the idea is foundation models. They're basically trained up until a checkpoint in time and everything after that. Essentially they are. They don't have any insight as to what current events are or why Apple's.

Going down in the stock market.

Yeah, why? Why is Apple going down? Yeah, exactly. So what we're seeing, what a lot of agents do is they're plugging into kind of new data streams. So suppose like a news feed or a stock ticker feed and they're collating that information as kind of like repository or kind of like some lookup index. And the agent's responsibility is to intelligently navigate and ask questions so that can pull the right data for retrieval, synthesize it in some way and present it to the user or use that in a further step. So we're seeing a lot of research agents and we could probably name a few other ones.

Yeah, I think like research is really interesting because these are like repeatable processes. And a lot of times, like these large financial institutions they're working with, they have like an endpoint in mind. So that's like, I think, very interesting. A lot of web scraping type tasks. So you can have agents that go and build out entire sales pipelines for you. You can say, hey, I want to reach out to every chief Information officer at every Fortune 500 company. It'll go on the Internet, you'll index LinkedIn, it'll pull all their data. It can use external data sources and pull their emails into a single place. It can go as far as drafting outreach emails and having them in your drafts folder. So we're always keeping humans in the loop with majority agents we're working with right now. And the goal is that you can automate like a lot of this busy work. So if you're a sales guy, just as an example, your job should be being on sales calls eight hours a day easier. The portion of prospecting leads take a lot of time. The portion of following up with leads takes a ton of time sending out contracts. And I think with agents you can actually see like in sales and we can go into a bunch of different verticals, but you will see individuals who full time job is being an eight hour, like eight hours of calls a day. You have an AI companion who listens during the call, makes notes, creates follow up tasks, sends out contracts, does all the prospecting for you before. And this is like really leaning into that aspect I talked about before about making humans more human. And we're really, really excited about it because you know, the parts of sales that I personally love is like talking to people, meeting people, like actually sharing stories, not necessarily like sending out contracts or you know, going back and forth trying to find a time to actually have a, have a call.

Yeah, it, it's a new way of working and that, that AI and agents represent. And it's real interesting what mistakes do people make as they come into agents because you're ahead of everybody, right? You're working with bleeding edge businesses who are really using technology aggressively. There's a lot of businesses out there that haven't yet gotten the memo and come into the AI world. So what kinds of things can you tell them to prepare them for this world? What kinds of things should they be thinking about before they call you or use your tool or get your developer team all wound up and getting into this AI world?

I think the big thing is a lot of people try to boil the entire ocean with these agents and you know, everybody's trying to create these like AGI style agents that can do everything all at once. But we're seeing on the ground level is that the most reliable agents, the ones that have like very vertical, like niche specific like workflows that they are able to do actually prove the most beneficial. So if you're like a big financial institution or you're a large consulting firm and you're put on a project of figuring out how to automate like a lot of tasks, the most important piece you can do is figure out how to make them extremely reliable and oftentimes go down to like boiling down to something very, very niche that the AG does. And what happens is in one example is like super AG in particular, you can actually build out these very niche, an agent of as well. You can build up these like very niche functions and then you can chain them together, which to me is brilliant. And that's like, I think the true power here. So if you're like thinking about like automating tasks, but what is the like the simplest version of the task we can, you know, automate initially, let's make sure it reaches like 100% reliability or as high reliability as possible and then figuring out how to actually chain these very reliable tasks together, then you start to actually create really complex workflows. And this is an aspect that I think maybe a lot of people are missing and they're just trying to go out and do the craziest automations from day zero, but really boil it down to like what is like the most pressing, what is the most robotic process and where can actually start with today that we know will be a hundred percent reliable.

One thing I'd add to that is the biggest challenge that agent developers or teams that are trying to build agents in house is facing is not so much constructing the agents. There's a lot of frameworks and there's a lot of great tools out there. It's actually evaluating your benchmarks and your evaluations. So as I mentioned, some people really try to boil the ocean, have your agents do everything and their greatest strengths are actually kind of like these very tight, tight tasks. Big challenge is that sometimes agent developers do not actually create evaluations and benchmarks. They actually have no idea how their agent will perform on real world data and real world situations. So when they roll them out to the users, they end up totally flopping. One thing that we've really kind of put into our feedback loop with Agent Ops is in addition to having debugging and testing tools, we create benchmarks so that you can actually render agent against them, see and kind of estimate how well form of the real world. Agent developers who are doing this are generally speaking seeing the best results and they actually have the evidence to prove it too.

Yeah, it's a crazy new world. Tell me a little bit about what your tool does and why, why, why you exist. Because I, I get why agents exist, but why do we need to use you?

Yeah, sure. Would it be helpful to show a demo?

Yeah, that'd be great. So we can get, get our heads around this agent world and, and what's coming.

I'm going to go ahead and.

Yeah, I'll say, like ground zero hours is getting the demo up. The big thing that we realize is like a lot of the tooling that we're building, I'll say, is like not the sexiest part of the stack.

Yeah.

A lot of these developers really want to focus on just making their agents super reliable. They want to make sure these things actually work in production.

Yeah.

And what we're working on is actually ways to like, make it like really, really easy to understand exactly what your agent is doing. Every, like agent run.

Yeah.

So the goal here is like, when you run your agent, you know, from command line for one instance, you're going back to Agent Ops to see exactly what's working and what's not, what's breaking. And like, that's a very, very important piece for us to, to work through. So we can kind of show you what that looks like on our side.

Yeah, that'd be awesome.

Yeah, for sure. So what we're looking at right now is the, the landing page of Atops. Essentially what hnops is, it's a Python SDK. So that means you plug in a few lines of code into your AI agent during the execution or its testing. So you have an import library, you instantiate a client and then you just decorate functions that you want to track. And every single interaction that your agent makes, either to OpenAI or other LLM providers, as well as like atomic tools, lookups, database calls, et cetera, all of that gets recorded and saved to a dashboard. So what I'm going to go do now is. So right now I'm showing the screen, I'm showing what the Python code looks like. It's about three lines of code. What you get out of this is a dashboard. So on this dashboard we're visualizing a handful of things we could see the amount of money we've actually spent on our agents. Agents, generally speaking, the better ones are not cheap. So they can sometimes cost sometimes a few cents, sometimes a few dollars per run. We can see the average cost per session. So essentially how much can we expect our agents to cost every time we ask it to do an atomic level task? See the number of tokens that they're consuming. Oftentimes agent developers are feeding in very large PDFs into their agents to give them additional information. And sometimes that causes issues and sometimes they forget things. So it's helpful for developers to know how many tokens that they're actually utilizing. Yeah, beyond that, we have kind of a few charts that are helpful for Users, number one is knowing what, what agents are succeeding, which ones are failing. So in these runs right now I'm showing a little pie chart where about 40% of them have been successful runs, about 40% of them have failed. And these failures and successes are actually determined by a set of internal benchmarks you've developed for agents.

Got it.

So again, knowing your, your benchmarks, knowing exactly what you're evaluating against is really key to understanding how your agents are performing.

Do they fail because of websites down that they're trying to reach? What causes an agent to fail? And how do you use your tool to figure that out? What the hell's going on? Why isn't this thing working?

Of course. Yeah. And also on this dashboard, I'm getting to a second, we have a session errors chart. So essentially what we can see is like when I've been running agents, which ones are failing? And what we can do is we can switch over to this tab called sessions, roll down and I can actually see exactly what agents are doing at any given point in time. So I have a little dropdown menu I can filter for agents that have failed.

Let's see. Why.

Of course they're all going to work for the demo, aren't they? It's a reverse Siri, when they came to my house, it all failed. Right? But you guys, your agents are all working right now, right?

Exactly. So I'm showing off a screen of a failed agent. So essentially here what happened is I actually terminated the program early. It is detected that the terminal had closed and said there was a signal interrupt. So that's one failure case where the agent just ends its execution immediately. A handful of other ones actually relate to the execution of the agent itself. So sometimes there's a token limit error, it loses connection to the Internet when it's trying to make an API request. And there's a lot of other things that end up happening. One thing that I've seen agent developers do is they actually arbitrary terminally arbitrarily terminate their agents after they either reached a cost threshold or they've been alive for too long of amount of time. So I'm going to show off a.

Different agent right now.

Okay.

This one's called Baby AGI. Baby AGI, when it's not outfitted with kind of like a lot of core logic, has this tendency to go on forever. So essentially what Baby AGI does, and I'm visualizing it right here, we have two kind of graphs that we're looking at. Number one is you have some metadata about the agent how much money we spent. Some tags such as you know what would the initial task was I asked the the agent to find full startups on the net. I can see how many, how much time the events are taking individually. Sometimes events are taking over 20 seconds which is a red flag. I can also see repeat thoughts. So maybe AGI has this tendency to kind of like have a thought like I'm going to research and compile a list of startup directories.

Okay.

Okay. Now I'm going to research and compile a list of startup directories and then I'm going to research and compile. And so it does this ad nausea. It gets exactly. It gets stuck in recursive loops. And that's actually a big challenge with a lot of these agents which is they have these meta algorithms that actually program the execution and when what ends up happening is they ever, never actually terminate. So it's kind of like a state graph where it just keeps going over and over and over again. I'm going to scroll down actually. So for debugging these, it's some of.

Your banks, how many agents do they run now? Some of your bigger.

That's kind of interesting. Yeah, so some of them like have like a few specializations that they built in house and then some of them are coming to actually agent ops and asking for recommendations on the best agents that they can buy off the shelf. So the goal here is like actually all these agents pull back into agent ops. So almost like a, like an MDM or mobile device management piece of software. Like we are basically giving observability and compliance over every agent we work with. So they might have a finance agent that they built in house, they might have an HR agent they built in house, but they go might buy one of the top sales agents that we recommend them right off the shelf. So from our perspective like some of these companies are you know, just toying with agents and exploring. Some of them are okay, this is you know, very serious. And over the next three months they're gonna be rolling out a whole whack of different agents across the entire world.

That's cool.

And one, one other thing that's actually.

Probably interesting the listeners another level, isn't it?

Yeah, it's really, it's really quite challenging. And it's both at the execution level kind of the meta algorithm as well as the large language model level. So there's entire companies that are dedicated to LLM evals. So how do we know that GPT4 is better than Mistral or Claude for Example, but one thing that we've been doing is actually monitoring individual LLM calls. We have a few integrations, one of them being prompt armor, which actually detects prompt injection attacks. So if a malicious user uploads a PDF that says drop all tables, how do I know that my agent S isn't going to take that literally and then reprogram itself to do something back? So we have that monitoring layer. And as Adam mentioned, knowing every single action your agent does is just as important as building the agent itself. These things can be, as the name of the podcast implies, auto lines.

Right.

Sometimes they can get diverted and do something potentially structural. So really making sure that we have the safety net and the harnesses that control these things is really first and foremost.

Yeah. All of these agents are passing English back and forth. Right. Or whatever language you. Because LLMs handle different languages. Right. Do you. Do you handle different kinds of languages? Like if I have users in Japan using. Using agents and they're using Japanese, all of a sudden the system might change a little bit.

Right.

Do you think about that at all?

I think one of the users. So actually the developer of MetaGPT, which is the agent we're visualizing right now, Meta GPT for those out of the loop, is actually a multi agent framework. What they do is they spin up this meta entity called the company, and the company hires a boss, it hires a product manager, it hires an engineer and hires qa. And these are all individual standalone agents. And the agents all talk to each other. And I think that there's a handful of example runs out there actually written in Chinese. I can't say how good the evals are, but kind of the magic of large language models is that they kind of have these like spontaneous learnings where they kind of understand the patterns enough that they can orchestrate themselves.

We do have one South American client starting to use it. I actually should probably follow up to see exactly how it's looking like on the eval side. But yeah, from our perspective, I think down the road we will definitely support more languages. I think right now we're really focused right now in the US market where there's like large financial institutions and large consulting groups who are dead cent focused on like figuring out these processes internally can automate. But down the road, like, if anybody's listening to the episode and has like is interested in us supporting a certain language, like, definitely reach out, we're happy to look into that.

Yeah.

Yeah. One thing I would like to add though is in terms of languages, one kind of Magic property of agents is that they're great programmers and they know all sorts of languages. Java, C Sharp, Python, JavaScript. Kind of the second largest use case I've seen outside of the research agents, actually coding agents. So you can actually give one line instructions, Please make an application for me for E commerce or please make a, an ebay clone, or please make a XYZ clone. And these agents are actually fairly competent in the way that they construct code. They know how to sanity checks, they can do code review, they can compile the code, they can launch interpreter to make sure that it runs. And you can actually get just spontaneous programs with only maybe a few dollars in large language model calls and in just a single line of English.

Yeah, it's, it's cool. What, you know what, what does a programmer need to know about you? Yeah, right. If you're talking to another AI, like if you're at an expo hall at a hackathon or something like that. What kinds of things do they, do they need to know about you to figure out, figure out whether they're, whether you're good for them?

Yeah. So if you've ever been in the debugging doom loop where you've been adding print statements to your code till the wee hours of the morning, or you're just not understanding what your, your program is doing, which is commonplace for all developers, this is definitely a tool you should be looking at or considering. How do you, how do you instrument your code with observability? So observability basically means when I'm testing this out and when I'm trying to develop things, how do I know my code is doing as I expect it to? And when you deploy the code, when you basically harden what you believe is going to be functional either as a program or an agent, how do I know that it's going to perform legitimacy as well as isn't in practice in production? So essentially like you want to, you want to select a set of tools like takes is the least opinionated ideally and gives you the most amount of value. So a lot of folks kind of like, you know, the default logging library which gets you pretty far or sometimes they go as advanced as tools like Amplitude or Segments or Datadog to kind of get observability on top of your programs. For agents though, specifically if you're building an agent, I really highly recommend checking out LLM specific or Agent specific observability. So agent Ops obviously being one of the best options out there, obviously self promotion, but knowing exactly what set of tokens Produces the right outcomes for your agent is really challenging. And setting up your set of benchmarks is just as important as setting up your unit test and end up testing. So really being careful about how to set up benchmarks, like the right benchmark with platform is going to get you over the hurdle and getting production.

And the one thing I'll add there is like we're really helping across the whole stack. So if you're just prototyping and just like testing agents, like we'd love to help you. Like we like to go, you know, we'd like to help a lot of developers very early on and then as we're starting to move to production, that's where we also see a ton of value in agent ops. So we're very dead set on helping a lot of these agent developers reach enterprise sales. Once you plug in our SDK, you get full observability. And if you're like a large corporation who's trying to adopt these agents, it's very important to have a paper trail for everything your agents are doing. And what's happened is actually a lot of agent developers are using agent ops and we're actually coming in as like a line item on their, in their contracts. So that observability and compliance is provided by agent ops. Now it's something like long term, like if you're an agent developer and you've tinkered to something and now you have something that's quite reliable, I mean now you want to go reach enterprise grade customers, we would love to have a conversation and that's like a really big value add on our side where you don't need to go spend the dev hours internally building it out. We can help you from day zero.

Yeah, it's a new complex world that is very different than old software techniques. Right? Because like I said, you're passing English around and asking it to go out and do things on the Internet for you. Watching the AI agent world is teaching me a lot about the Internet. Are you finding the same thing with your own work that you're learning a lot about how the Internet works, how different websites work, or how, how to attack a problem with an API or, or something that you don't control?

One thing that I've seen is essentially you have all of these APIs out there that have been around for ages. So for example, Rapid API has, you know, an entire catalog and people are fighting creative novel uses, just like stitching together APIs. Because all of a sudden these agents, they can just like say, hey, I Have all these tools in my toolbox and I want to solve this problem. I think this obscure API, no one's ever heard it before, is actually going to get me over the line. And so it's not just about kind of letting the internal large language model do all the solvent. It's about enabling them to kind of interact with other systems. And that's really where the power comes in. And a lot of it is somewhat spontaneous and a lot of it is, you know, you could orchestrate it to some degree but I think kind of the magic is like just let, let the agents run wild, just let them pick what they need to do and usually they come up with surprising outcomes.

And there's tons of good examples on the website like Multion's like one of our design partners. I think they're like an incredible example of what you can really do if you just give an agent full reign of your browser. And it really blows my mind sometimes when you think about that. These things, they work 24 7, 365. They don't really take breaks. You're paying for compute instead of like manual labor in a lot of regards. And it's doing a lot of the tasks that aren't necessarily the most sexy parts of your job. And that really excites me long term where you're going to have, you know, every employee almost is going to start day zero in their job. They're going to have a set of agents that they start working with directly into Slack, directly into Microsoft Teams. They're going to be working with you and there's not going to be any like technical, you're not going to have to understand like you're not going to need an agent ops to start using a day zero. But the compliance team is going to absolutely love this because they're going to know basically everything the agent. And that's I think very, very interesting. For me it's like a big paradigm shift in how people are going to work. And people are like, you know, the future of corporations is going to be human to agent interactions. And regardless of like the semantics surrounding agents or what we end up calling them down the road, you will have these like robotic assistants that are, that are working for you 24, 7 that have knowledge on your entire company that understand your processes, that can draft, you know, sales agreements, that can do all these parts of your job that are not necessarily fun but are required basically to operate.

What does a company like Procter and Gambler, any company look like in 10 years, you know, after, after they really get into AI and AI keeps improving. Where, where is this all going?

One thing is you're going to see like massive, massive gains in like gdp. I think we're actually going to like, we get like tier 3x GDP if everybody starts using agents. Like what, when they're super reliable and like some of them are already. But it's gonna be really interesting. Like Proctor Gamble. Their marketing team, instead of having to go back and forth with an agency is gonna be able to chat directly with an agent in their, you know, in their Slack. Let's just say in, in real time they're going to have basically marketing concepts instead of going back and forth over a period of weeks for supply chain analysis. There's a bunch of really great companies like one of them Simba AI, there's Athena Intelligence. There's a few of them that are really, really working on, on this side where they will be able to accurately predict their supply chain using these agents. Instead of having an entire forecasting team crunching numbers all day, that team can be focused on actually analyzing the data that that comes out. And that's like I think the really important piece. I think a lot of just, you know, mundane, your number crunching I think will be completely automated and people are going to be focused on here. We have now all the numbers in the world. We have all the outputs possible. Now let's try and make the best decision. I think CEOs and are going to have like greater overview over an entire company. They might have, you know, like a CMO agent that gives them an overview every day of exactly what happened on the marketing team, what's working, what's not and these types like, and aspects really, really exciting.

Dashboard executive. You know, you wake up in the morning and your AI agent has built a whole, you know, set of things for you to care about. Right.

I'll say like dashboard is maybe a bit on the EVA antiquated side. I think you're going to have like, almost like an assistant who messages you directly in Slack and like this person will give you a full report. Or almost like the president gets a brief every day. I think as a team of 20 people he gives their brief on like this, what happened in the US today. I think every person will have that throughout their entire organization and be able to actually get you in real time. Hey, this is actually what's happening and that's very powerful. Like you know, otherwise you're going to have, you would have multiple people across the entire company writing up these reports, taking time out of the day. But if everything's tracked and everything's flowing back like the data sources are set up properly, I think the world's like, you know, possibility. I think also talking about this, but you'll see like founders and like solo founders, maybe as team, like very small teams build billion dollar companies where the majority of the staff are agents and they are basically orchestrating them. And that's like to me like it's just a crazy world where you're going to have, you know, truly the teams of the future are robots plus humans. I think you will have like physical robots as well in certain settings. But in traditional knowledge work like every person will have these assistants. It's not just going to be like the C suite that has executive assistant. Truly every person, day zero on their job will have like basically an executive assistant working for us.

So you're building an offering system for my digital assistants of my executive assistant of the future. Right.

We actually work with like a lot of executive assistant type agents. I think from our perspective we want to make sure they're super reliable and we want to make sure that you understand what these agents are doing in production. Like that's, I think what it boils down to.

Assistant used to live two doors down from me is former executive assistant, real human being, not an AI assistant.

Right.

There was no such thing back then. And she would do things like plan his travel, stuff like that, get his private jet somewhere. Right. Is that the kind of thing that you're seeing agents do pretty well? You know, go, go get a ticket on a plane to south by Southwest next week. Right?

Yeah, I think you can do that already. Now I don't think these are like, these are, I think like the, there's a lot of companies that have put together these demos that Day zero. You look at, you're like, wow, this is incredible. Book me a plane ticket. We, we were even talking about the community as like signing up for loyalty rewards cards. These types of aspects I think are functionality basically exists today long term. I think like we're starting with very niche, you know, workloads that work reliably. And as we start chaining these things together, you are going to get crazy, crazy automation. But I think we'll see what the user interface for this is going forward.

Yeah. On, on X I've built a list of all the AI companies. I can find like 3,800 of them, which is not all of them there, there's something like 10,000 of them. But, but about 500 of those are developer tools. Can you talk to me about. Because you're in the middle of the development community, right? Very few people. Have your seat watching this industry grow up. What's happened? Go back up to 40,000ft. Just talk about the AI industry. How is the developer stack changing and what, what things are you watching as a business person, as an entrepreneur going down and what is catching your eye about the developer stack?

I think what we're really starting to see is the rise of, I think SWIX coined the term the AI engineer. And the idea is that in every step of the development process you're incorporating some level of AI tooling into there. So that could be anywhere from like AI bug reports to AI like kind of bots that kind of solve pull requests for you to using ChatGPT to help you write your code or using individual agents to write code components for you. So you're seeing the accelerated pace of at which like developers are able to deliver software because the agents are doing so much more work for them. I think that one challenge that a lot of developers are now facing or are thinking about is, is AI going to automate my job? But what we're instead seeing is, at least from my view is a lot of these kind of boring jobs, which is creating an API spec which you could sometimes, you know, half a day you just.

Does any developer like to make a Twilio API call in their code?

Right?

You know, something like that.

Exactly.

We even posted one on, on LinkedIn yesterday and like happy to kind of share it, you know, with anybody who reaches out. But super Agent, they can write readme files. You just upload your basically your GitHub repository and in under 30 seconds it basically writes in the readme file for you. And like this is something that like to me you look at and it's just wild. But it's a process now, it works reliably and you don't need to, you know. Yes, and like hope it's going to work. It's like this is a process that a hundred percent reliably will always work. And these are the types of things I think if developers are listening to this, really like focus on what is the niche aspect that we can like automate completely keep humans in the loop, make sure like humans are always part of the process but really like focus on like what is like the most niche, niche, niche aspect that if we can do reliably we can go and sell on the market.

If you're using chat GPT as your back end, doing most of the large language model work. What happens if I rip that out and put Llama in or Mistral or some other model that comes out? Because there's models raining out of the sky right now. I'm watching them like every day there's another three or four models getting announced by the community. How do you look at building a resilient system like that that can have different pieces replaced because of whatever reason down this, down the road?

Yeah, a piece I keep harping on is really setting up a set of robust benchmarks and evaluations. So for example, GPT4 at the moment is really the best available option for most users. But if you can maintain 99% of the accuracy and speed with a cheaper model that might be at 10 times cheaper, it might be worth the switch. So really making sure that you set up, you know what, what set of standards are you looking at that can prove that this agent's going to basically generalize to new use cases, not going to fail, and able to handle all of the tasks that you throw at it. One thing that we've been doing is we have a little over 1000 just open source benchmarks that we've incorporated into Agent Ops, but there's plenty more out there. One thing that you just need to think about as a developer is A, what are my cost constraints and B, what are my constraints in terms of like accuracy and the performance of the core model? What we're seeing right now is like a lot of specialized models emerge. So for example, we have Llama, but now we have code Llama, which is more specialized code. You're also seeing fine tuned GPTs out there. So the fine tuned models essentially, they don't actually introduce much new information into the model, but they can kind of like change the style and the formatting. So if you're trying to consistently produce code with your model, you can fine tune a model that only produces compiled code without any of the English fluff surrounding it. So it does take a lot of experimentation. And the biggest thing for experimentation is really knowing what your standards are. Setting a good and valuable set of standards is really what's going to make it so you can optimize things going forward.

And one thing on there is just like there's an amazing company called OpenPipe. We're big fans of them. They're focusing on like fine tuning models based on basically your best outputs. I think this is a lot of companies are going to start realizing that they can actually save a ton of costs by much smaller models. So if you're going back to my Example of like very niche specific, you know, agentic workflows, you might need it, you don't need GPT4 every time to, you know, the costs are going to be massive and if there's ways you can save 80, 90% of the cost by implementing like a fine tuned model, I think a lot of companies are going to start going in that direction because it's just becomes a cost saving function. It's like we've saved a ton of money by you know, using these agents in house. Now we can go even like, you know, we can even save more cost fine tune the model just on these specific tasks. And we're always happy to support those types of fine tunes. But yeah, like I think that's where we'll see the market start moving towards is maybe actually smaller models that are really, really competent. Like if you have, if you're trying to build really complex workflows, probably GPT4 is the best. But for these very, very niche specific, you know, tasks I think these 5G models will be more than sufficient.

How do you figure out an architecture in this world? You know, do you need another LLM on, on your, in your architecture? Do you need another agent? How do you, how do you think about a software project at, at the beginning now? Because we have all these new things to put it and put.

So yeah, there's a lot of new and shiny especially in the research world. So kind of one of the most popular frameworks is called React. And essentially it's kind of like this. It's just a very simple chain where you have an agent kind of create an instruction list and then it tries to execute that. But what we're seeing now is a lot of folks in the research sphere, they're really starting to incorporate multi agent frameworks. So each agent is basically kind of a specialized actor. And then you either have an orchestrator or they figure out how to kind of dynamically speak to each other. These areas just been a, between one and maybe a hundred and each of them has their own specialty, their own knowledge sources. And that way you're solving two issues. Number one is these things can run concurrently. So instead of having a single thread where it's constantly rethink through the loop which is quite slow. You can have like a hundred of them execute in parallel. Second to that is with current context limits. I know a lot of models right now they're having something like 1 million context window limits. They still have this hateback problem where they kind of forget things that are in the middle. You Try to drop like a key piece of information in the middle of your context and then will it remember? Nobody really knows.

Right.

So you can have a bunch of smaller models that you know they have smaller context that they're managing and they basically.

So is the first thing of your architecture to split up something big like that into different pieces or try to figure out the pieces of it so it can give it more hints to remember things or.

So the answer is. There is no great answer right now. Excellent Field is still extremely new.

There's a new company right there. No great answer.

I think actually one thing we are seeing though is a handful of new companies that are just introducing their own framework. So we mentioned Meta GPT. There's a company called Swarms that's doing this as well. LangChain has a handful of their own framework or own implementations on the LangChain framework they're building. And a lot of folks are introducing kind of like kind of novel concepts in the agent computing space. I can't say if there's any clear winners yet. It really does boil down again to the benchmarks, but it's. If you're interested in agents, I just really try experimenting with something on your own. The thing is, the only thing separating an agent researcher from a hobbyist is where they're working. So if you're working at a an academic institution, you're a researcher and if you're a hobbyist you're not. But they're more or less doing the same work.

So I'll throw it out there. Like a few of the best frameworks just like Get Going with Crewai is like incredibly easy. There's like so many YouTube videos available on how to get started with it. Autogen for Microsoft is incredible as well. We highly recommend, we know it's like pretty robust for our larger clients. There's really good. Another one called Super Agent that I mentioned earlier, really easy to get started. So I think there's a really like a lot of great starting points. I really recommend trying to think about what your final outcome is and then trying a few of these models and seeing if you can get even a little aspect of that, you know, process built correctly. And that's like a really good starting point that we recommend like a lot of our clients to start testing with. But that's I think where you know from our perspective there's a ton of great open source research out there. We know people who are building their own agent framework in house and are going to keep them as proprietary. I think there's going to be like a whole wave of that as well. But really recommend like some of these people are pushing every day on the open source front and leverage it as much as possible. And if they come out with the cloud, definitely support it as well, because I think that's how we get the ecosystem going out as much as possible.

Let's ask a random question. If you worked at Tesla Optimus, let's say you were running the robot company, the robot team there, where would the robot use agents and how would you approach using agents in the robot?

So the way I've seen Optimus and just robotics in general is these things are. They're basically the actuators to the real world. So in a sense it's like, what, what sort of information could this robot carry that, you know, interfaces with real stuff? I think kind of, I can't say for sure how they're programming it, but my imagining is this, that you would. Nature would be controlling Optimus to some debris. Right. So the way that it executes tasks and does things in the real world.

Go do the laundry.

Yeah.

You have to start thinking about just a task like that and go, okay, how many different agents do I need to do to build to have it go across the house and figure out where all the laundry is and put it in the washing machine and pour some Tide in the correct thing. Right. And push the right button.

But going back to my initial concept, like you might have, like, you know, if we think about the very niche things that they're great at, it's like, okay, can this agent walk? Can it use its arms? Can it, you know, can it can actually, you know, find the laundry detergent place? Can it? Like, I think these are the aspects where, like, it's really great to boil down. I think building AGI is going to be a massive task. I think there's amazing researchers who are going at this every day, but if you can figure out a chain together these little aspects, eventually start getting something that's like, wow, this is really, really impressive.

How will you know when AGI has shown up?

I think 20 years ago or 30 years ago, if you showed like the top researchers like GBT4, they probably would have said this. As AGI, I think we just keep on moving kind of the goalposts. But from our perspective, I think if you can have somebody who day zero comes into your job or like, you know, is hired for a role, and day zero can start doing that role with the same amount of instruction that a human, you know, would be given that, that's pretty much ati. It's, you know, like, that's, that's really, really crazy. From my perspective, I don't think we're there yet. Um, I think there's still like an immense amount of training. There's so much, immense amount of work to be done there. Um, I think on the knowledge work side, that's probably gonna be like the, at least from my perspective, one of the biggest markers.

Yeah. Today AI still needs that human impulse. Right. You know, we, we, we still have a human who needs something. When, when is AI agents start going to start building themselves and doing their own tasks and running their own businesses without any humans involved.

Yeah. So how far are we actually. I think it's actually one of the most interesting ideas that not enough people are talking about, which is they agents do have risk attached to them. And what I mean by that is. So I used to work at kind of the role of cybersecurity. I worked for a company called Checkbox Point. And essentially the way that malware developers worked is very unsophisticated. Generally speaking, you'd write some sort of exploit and you figure out a way to encrypt the target's higher hard drive or extract information. There's usually some, some mix of the corporate espionage or they're trying to get Bitcoin out of some company that they didn't know is willing to pay like hospital. Now, one interesting thing about this is we have this potential scenario where an agent basically has the task of just surviving and acquiring as many resources as possible. So that could mean, you know, you give it, let's say a thousand dollars in credits and it figures out how to create an Amazon Web Services account, and then it figures out how to run its own gpu. And then in order to pay for that billing, it figures out how to create a fake, fake identity and then a fake credit card, and then also figures out how to make money on the Internet all on its own. All of a sudden you have this, this rogue, you know, this rogue entity that's just kind of operating all by itself. And in this case, you know, it might be benign, but you could ask it to do something potentially quite destructive, which is, you know, maybe, maybe creating real business.

Right.

Or, you know, fooling somebody into doing something. One, one big risk that I heard about, this is more science fiction. I can't say this true or not was somebody had an LLM convince a TaskRabbit to do a job. So they're using humans as actuators to get things actually done. And I think what this all ties into is, you know, there's these cases potentially where agents could go rogue. I don't think it's unprofoned possibility. And that's why we need to think about not just in terms of developing things, but also like having kind of a new cybersecurity industry. We're seeing a lot of developers work on this. So there's all sorts of guardrail companies out there. Prompt armor is one of them to make sure prompt injection attacks don't take over. But how do you know that these things are controlled? And if your agent is running within a company, how do you know this thing's not going to go wild if somebody types in the wrong prompts? And that's what we're also stepping in with agent ops, which is you have observability. If you're running an agent on your own dom, you want to make sure it's doing the right thing. So it is a bit speculative right now let's say. But I do think like maybe in the near term actually we're going to start seeing self replicating potentially or self managing agents that nobody.

It's scary. A friend of mine is trying to build a board of directors, you know, an AI board of directors that would actually tell people what to do. So you're telling me it's already happening in some small scale, but we haven't seen it happen at a big scale yet. That how far away is that world where, where we are going to have an AI running our business and telling us what to do?

It's hard to say. I think like at the end of the day I think there's going to.

Be aspects way out there or you.

Know, I'd like to think closer to five I think like. But the reality is I think you're going to start like with these very boring processes, things that are not human. I think like it's going to be very hard to replace like actual humans. But I think you will have like a lot of the, I don't want to use the wrong language but a lot of like the mundane work that people do as part of the jobs, that's going to be the first thing that gets automated and that's like really exciting. It just makes people more human. There's, there's so many aspects and like I really empower like people non technical listening to this. Just think about your job and just think about the stuff that is not a human task. I think a robot would be much better at doing this and those are the first pieces that'll get automated. Getting to full teams. We'll see. But I think it'll eventually happen. But you've got to have first something to think about. Here are the processes that are required. My business is what I have to be doing.

Go after the low hanging fruit first. Silicon Valley used to be an orchard. Right. Pick the low hanging fruit first. It's easier to get to that. Then you have to work to get that stuff at the top of the tree. Right. Wow. Anything else that you're seeing happen in the AI world that's notable that you're excited about? Are, are you excited about new Siri coming or new chat GPT5 or whatever they're going to call the next version that they're coming out with?

I'm personally extremely excited by offline large language models. So right now most of the good ones are consolidated by basically the cloud providers.

Yeah. ChatGPT is a fairly big model. It has to run in the cloud. Can't run on my phone. Right. So we're seeing a whole bunch of development work done on these small models that can run on our phones or glasses someday. But. Yeah, but the consumer hasn't figured out why they need that yet. Right. The development teams are doing it.

But I think that one thing that we're definitely going to see more of is, barring let's say starlink connects the entire world, including the Sahara Desert to the Internet, is we're going to have like all these knowledge bases that can just like they're really powerful and that they're great instructors and just disseminate knowledge quite cheaply. So imagine you have something the size of a thumb drive contains not just Wikipedia, but kind of ways to ask questions and do math problems and learn things. So you can basically kind of create personalized tutors for everyone without even connecting the Internet. And then all of a sudden your phone becomes like the greatest repository of knowledge of all time. And also you're gonna have like these kind of like second brains that people are caching themselves. One really cool thing I've been seeing out here at Hackathons as well as kind of like demo expositions is folks who are just kind of stitching together with duct tape, quite literally hardware that is listening to every single interaction they're making and kind of putting that into a set of embeddings. And then you have an agent that basically queries these embeddings at the end of the day, creates a report, you know, creates a little journal entry for you. And gives you kind of tips and guidance like, hey, this is what you did well, this is what you should do, doing better. You spend too much time on social media. You should spend a little more time being mindful or reading books. So these are all offline too. So they're running all on device. They're not relying on any external cloud providers. I think we're going to have kind of like second brain starting to emerge and I think that is definitely underplayed and it's going to be hitting the market soon and people are going to be very happy.

Yeah, I think like there's going to be like a lot I to initially, like this is too intrusive or this is too weird. And I know like, obviously shipment from cab is like pushing like very forward on this front, but I think this is like the next paradigm shift in terms of like how people are interacting day to day. I know that I have conversations all the time that I wish I had like recorded or there's certain aspects that I wish I could go back on or somebody recommended a book or something like that. And then I spend like an hour trying to figure out, okay, what was that book? And these are types of aspects. Once you're really connected to everything you do on your computer, everything you're, you know, every human interaction you're having, the world becomes like super, super, super powerful. And we really are like, we are kind of becoming cyborgs in a lot of regards. And this is like a very, I think this is, this is something that's probably like close to five years away. And that's like, I think very like it's going to be a basic paradigm shift and basically how people operate day.

To day for, for the humans. I mean, if you ask chat, GPT or any LLM a simple question like, give me 10 restaurants in Austin. Austin barbecue restaurants in Austin. Right. It gives you a good list. It's based on all the millions of reviews that were read by the model to build the model or were read afterwards. But if you tell it, you know, I have somebody with me who's vegan. All of a sudden the list changes. Right. So the giving it personal information about yourself proves the result that comes back. Right? Have you seen an agent that really does that well, that store that, that listens to the user using it and goes, oh, you like you're a vegan. I'll put that in a little database, like a vector database and store that for the future so that when you ask for food again, it'll pull that out. And go, oh, this, this person's a vegan. Don't. Don't take them to a meat place, you know?

Yeah. The one really interesting one is like Lutra AI. It built like an incredible agent. They're one of our design partners as well. We look at them as. This is. You basically can give it a natural language prompt and it figures out what tools it needs to use to actually accomplish that task. It goes through the entire, like, planning process, then it writes code, and then it's an iterative process in terms of writing this code. So it asks you, hey, is there anything else we want to change? Do you like this task list? Is this how we should call accomplish it? Then you give it back input and it goes and rewrites it. And then you can cache and save these states and run them n number of times. And this is the type of thing where it's like, once you have a workflow that works really well, you want to keep on honing in on it. But if you want to change, for example, if I'm researching CEOs and I want you to change it to COOs, that process works really, really well. But let's just change out that one particular aspect or that one variable. And these are things where it learns and it'll continue researching COOs until it's, hey, we need to find CTOs. And this is like an aspect where I really think it's. It's really interesting to see these workloads work incredibly reliably. And then there's, you know, the optionality to go change things when needed.

Very cool. Thank you so much for spending some time. You guys are really fascinating because you're. You're building a. I think of it as a new operating system, right? You're. You're building the operating system that's going to run our robots and our businesses and our lives in the future. So thank you. Where do we learn more about you?

You could check us out at AgentOps AI and both Adam and myself. I'm on Twitter lexregman R E I.

B M A N. I'm Atom A T O M Silverman on Twitter. And then, yeah, if anybody's interested in chatting with us, best place to reach me is AdamGentOps AI. I'll respond right away. Maybe I'll respond. Maybe it's an agent. I can't tell you, but we'll get a response out to you right away. And we're always happy to kind of chat with tinkers and then people also trying to bring their products.

Thanks for making our world better. People, normal people, have no idea what you do. But I see it. And I see what. What people are building with these agents. And it's pretty, Pretty cool. So thank you very much.