What's an AI Agent? And Why's That Hard to Define? (The Agents Season, Episode 1)

0:00Hi, welcome to Linear Digressions. In the next few episodes, we're going to do something a little bit different from the usual linear digressions where each episode is its own standalone content. What I wanted to do was unpack AI agents. And the time seems right overdue even for an exploration of AI agents. But that is more than one episode's content, to say the least. It's more than two episodes content. And as I was thinking about it, there was a whole bunch of different facets of AI agents that I wanted to explore. And so I thought, rather than trying to bounce around and do that piece by piece, actually explore architecting a whole season around exploring AI agents from different facets and kind of build up a picture piece by piece about what are the most common and interesting aspects of working with AI agents right now.

0:59So that's what we're going to be starting with this episode and continuing for the next 10. So there's a lot that we're going to be covering. So if this is the first episode that you're listening to in this series, welcome, you're joining at the perfect point. If you are joining from one of the later episodes, you decided to rewind a little bit and start from the beginning, also welcome. And so without any further ado, let's get into it. You're listening to Linear Digressions. Right, so why agents and why now? This isn't an abstract philosophical question. This is something that's actually very timely.

1:37Our whole society is starting to figure out what it means to delegate something meaningfully to an AI. This has business implications, it has personal implications, it has social and economic and arguably ethical implications. But unpacking this from a technical perspective and digging into some of the underlying core themes, it's quite interesting. But first we need to start with a definition, which is, in my opinion, surprisingly difficult to pin down, which is what even is an AI agent? Here's the problem. AI agent is everywhere right now. If you are reading product announcements, if you are hearing people pitching VCs, if you are reading tech headlines, everything is AI agents.

2:24And this means a very wide range of things. It can be everything from a slightly smarter chatbot to something that is theoretically autonomously running an entire company while you're sleeping. And that ambiguity is really challenging. Some of it maybe is sloppy marketing, but I think it's reflecting something deeper, which is that this is actually a concept that's tricky to pin down. So before we start talking about all kinds of stuff that agents can do or the components of an agentic system, what it means to trust one, figured it was actually worth the exercise to work on definitions for a little while.

3:05So let me go through some examples because the abstract definition I think is not very useful. I think watching the concept emerge as you think through increasingly complex examples is a little bit cleaner. So with that, let me start with example one, the baseline. Let's start with ChatGPT. So you open it up and you ask a question like, what are some good podcast episodes about AI agents? And it'll give you a list and then you read it and then you go to iTunes or Spotify or whatever and start listening to one. Conversation is over. That's the end. So is this an agent? So it's very useful. It's impressive, especially if you remember that none of this existed a couple of years ago, but nothing has happened in the world.

3:53So you ask a question, you get a response back from the AI. And the only thing that's changed is that there's words that showed up on your screen. So the model isn't doing anything like accessing your emails or going into your calendar. It didn't even search the web, maybe, let's imagine in this thought experiment. You asked a question, it responded, and then the end. So this is a baseline. Let's call it a responder. The interaction between you and the LLM is complete when the answer is generated. There's no loop. There's no persistence. There's no consequence beyond just the conversation that you and the AI have.

4:26And that's an okay baseline, like first mental model of what AI is, which is, of course, not wrong. It just doesn't have the full picture of an AI agent. so let's start to layer in now to our second example tools and actions we'll add one thing same question but this time the model has access to a web search tool so you ask the same question about podcast episodes about ai agents and then it can run a web search say looking for recent podcast episodes pull some results incorporate those into its answer and then that's what it response with. You may be wondering, okay, but wait, isn't that an agent?

5:04It used a tool. It took an action. Yes, it did take an action. So it reached outside of itself and it did something to the world. Namely, it kicked off a web search. And that is genuinely different from the baseline. But there's a bunch of stuff that it didn't do. It searched once, it got the results, and it stopped. Again, there's no loop. There's no feedback cycle. It didn't look at the results and decide it needed to search again with a different query. Didn't notice that one of the results was broken and try a different source. So it executes one action in service of one response and then it's done.

5:38So the key question isn't whether a tool was used, it's whether the model is observing the results of the actions that it took and deciding what to do next. And it's that loop between taking an action, making an observation, deciding what to do, and then acting again. That's what's starting to look like agency. So we're getting closer in that we have a model with a tool that can take an action, but the autonomy isn't quite there yet. Let's go one step more advanced. I'm going to change the scenario now and say that instead of looking for podcast episodes, you are going to ask an AI to book you a flight to New York City next Thursday, cheapest option under$400, I'll see it if available.

6:23Oh, and you have some frequent flyer status on United. So if you can get something good from United, that's going to be your choice. So now with this example, think about all of the different component steps that are going to be required for that task to be carried out. So you have your element of searching. Again, it's going to have to go search for flights. More than likely, it's going to be looking more than once across different parameters. once it gets the results back it has to compare those options it has to notice when something doesn't meet your criteria and try again probably has to fill in a form has to handle a confirmation flow and at the end it has to know if it's done you may be thinking to yourself that's a lot of steps how does it know when to stop and that's actually one of the hard problems in agent design we'll come back to it in a later episode when we talk about planning but for now the point is start to notice what's different here.

7:16The agent isn't just responding to your message, it's pursuing a goal, and is doing that through a sequence of actions, and each of those actions is informed by the result of the last one. Now when we get into the research literature in some of these upcoming episodes, we're going to start to see concepts like act, observe, reason, act again, a cycle that looks like that. The model looks where it is, it decides what to do next, it does that thing, it sees what happens, and then repeats. So it's not executing this predetermined script. It's not an automation. It's a navigation. There's also another thing that's different here, somewhat by construct because of this example.

8:01But that is that there's consequences now to what the AI is doing. So when you book a flight, there is actually, or there should be, there's actually a ticket that is assigned to you. There's actually$400 that leaves your bank account. You are going to end up in New York City next Thursday or not. And if the agent makes a mistake and you end up not in New York City, or you have a middle seat, or any one of a thousand other mistakes that it could make, you're dealing with those consequences. And this is one of the things that's sort of most tricky about agents in real world scenarios is when they make those mistakes, can you catch them?

8:41Can you correct them? And some amount of the time, they're not going to be caught, they're not going to be corrected, these mistakes will happen in the real world. And who's accountable when that happens? Super tricky question, and it is not actually answered yet. But it's one of the reasons why I think understanding AI agents and the components that are working inside of them is an important baseline for actually engaging in some of those higher order conversations. Also, I would say we have definitely now started to move into the territory where we're dealing with an AI that's what most people would call an agent.

9:12This is what you probably imagine when you're thinking of an AI agent. And it's a fair picture, but we can get even another notch more complicated. So let's do it. Example four, OpenClaw. This was actually the example that motivated me to do this series in the first place. So what is OpenClaw? We're going to work our way through there in a couple of different pieces. But what are the types of things that you can do with an AI agent like that? It can be doing something like managing your email. Not that it's just reading email when you ask and responding to it according to a script. It's managing it.

9:48Like it's running in the background. Somebody sends you an email. The AI sees that and can respond to it without you ever being in the loop. Potentially. You might hope that it responds to it according to some criteria that you've given it. You may even occasionally find that it does something clever or unexpected. But imagine there's any one of a number of things that it can do in responding to emails, managing emails on your behalf. So when an AI agent is reading your email, drafting responses, and sending those, then we've got a really interesting version of this accountability question that I just introduced.

10:24Who authorized all of that? If an email gets sent on my behalf that says, sure, I would love to meet you for coffee next week. Was that me who sent that email? It came from my account. No, it wasn't me because it was the agent that did that on my behalf. And this is where it gets tricky. So what I might have done for that agent is I gave it a policy. I might have given it a set of conditions. Here's how to evaluate which coffee invitations I'm going to accept or politely decline, try to reschedule, whatever. And then I might have stepped back and the AI agent is just applying that policy, applying that judgment ever since.

11:03And that can happen without me being in the loop at all. This is also a setting where questions about the agent being done with its work, it's booked the ticket and you have what you need. Those also start to become less meaningful in this paradigm, because it's not working toward a single goal or a single task that you've given it. It's just managing this ongoing relationship with the world on your behalf. And so this is a fairly extreme example. It's one that we have now with some of these new agentic systems. So it's one that we need to be thinking about. But it starts to go pretty far into the extreme past just there's an AI that's interacting with the world on my behalf in a discrete sense.

11:46And it's a little bit less of a technical question and arguably at a certain point, like this philosophical question of like, what have I actually handed over to the AI? What have I delegated to it? It's a piece of judgment that it's exercising on my behalf. It's doing this decision making that used to require me to be present, but doesn't anymore. And it's pushing us in, I think, some really interesting and important ways. So by the time we get to the end of this series in a few episodes, there's actually an AI agent that I made. This is a little more in the vein of example three than example four.

12:19This is more like the plane ticket agent than the fully self-driving car of my email agent. But I actually have an agent that I use to help me with some of the day-to-day tasks of creating this podcast. I'll take you inside how I built it, what it does. There are some things that it does really well. There's other stuff that I have kept for myself. And I think that's actually pretty typical of agents in this stage right now, especially if they're agents that you've built for yourself, that a lot of the task is about thinking what even should go to the agent in the first place, what stays with you, and what lives in the middle region between them.

12:56But anyway, we'll get there in the fullness of time. But just as a little bit of a teaser, that it has been one of the fun things about bringing this podcast back, because it's given me an excuse, and to be totally honest, like, a true need to figure out some of these systems, because I don't think I could do this podcast in its current form without that little extra help from AI. So thank you to my AI agent who will be helping clean up this episode and reading the transcript and generating the newsletter and all of this sort of stuff. And of course that conversation will be featured in a few weeks.

13:30So with all of that, there are many different definitions of AI agents. In the course of research for this episode, I came up with several different ones. Russell and Norvig, these are some of the AI textbook writers, they have a definition that just says an agent perceives this environment through sensors and acts upon that environment through actuators. So that's one of the most established definitions maybe, but it doesn't really capture what we're doing here. Anthropic has a slightly different definition that emphasizes consequences that agents take actions in the world with real effects that persist.

14:08The definition that I'm going to use in the context of this conversation over the next few weeks comes out of the REACT paper. We're going to cover this in a couple of episodes. And in that paper, the authors emphasize that loop, observing, reasoning, and then acting. Observe what happened, reason about what to do next, and then act again. This is not just about a response. This is not even a sequence of actions. It's this cycle of perception and decision-making that keeps running until the goal is met. So that's the definition we're going to work with in this podcast. It's not because the other definitions are wrong.

14:44They're all catching something that's real. But because that loop is one of the things that makes agents really unique and interesting and hard. And it's running through everything we're going to cover this season. So when I say agent from here on out, that's what I mean. Something running that loop with real consequences on the other end. So with that, here's what's coming. In the next episode, we're going to talk about the transition from your chat GPT style chatbots to what we started to think of as real agents. This is when the models, the LLMs first learned to reach outside of themselves and learn tools.

15:25The papers that made this possible, what they actually showed, and we'll get into react properly and why that observe, reason, act loop looks so elegant on paper and is so interesting in practice. this. Zooming out a little bit over the course of the season, we're going to talk about some interesting facets like the failure modes of agents that you don't see in chatbots. How do you do evaluations of agents since they are so complex and tricky to evaluate? We'll talk a little bit about the economics of running them. And also I'm excited to dig in a little bit to multi-agent system when you have agents talking to each other and trying to coordinate divide and conquer work.

16:07As a little bit of a through line, we'll probably mention Open Claw a number of times over the course of this. But one of the things that's also been cool about this is watching the through line through some of the research programs as well. So as usual, for each of these episodes, we'll have some papers and some references that you can read on your own if you want to go back to the literature that you're exploring. So at the end of the day, what I hope you take away from this is that there's this underlying question of what does it actually mean to hand something off to an AI? Because that's the power of an agent.

16:42It can take real work off of your hands. It exercises judgment on your behalf and then has those consequences in the real world. When that happens well, it's magical. When it doesn't happen well, there's all kinds of crazy stuff that go wrong. We are all figuring this out together. let's go on the journey together. So with that, if you would like to see the output of my linear digressions agent in action, subscribe to the newsletter. You can find it on Substack. Go to substack.com and look for linear digressions. What you have in there is a recap of the notes for each week's episode, links to the source material, as well as some tips if you want to kind of get the distillation for how you might use what we're talking about in your day-to-day work.

17:29There's also a little section where I like to drop stuff that didn't have a place in the core episode, but that I thought was interesting along the way. So there's some unique content that you won't hear on the episodes. So if that sounds interesting to you, or you're just a little bit curious what some A plus AI agent generated linear digressions written copy looks like, head on over to Substack and look for linear digressions. So with that, a little teaser for next week. Next time we're going to be talking about React, that reasoning acting observation loop. We're going to talk about tool use for the first time.

18:06And in that context, we'll also name drop MCP. So it's going to be a good one. I will see you there next week.

18:17This has been Linear Digressions. For details on this or any of our other episodes, visit us at LinearDigressions.com. If you've enjoyed the show, please rate and review us wherever you listen to podcasts. It really helps new listeners find us. You can reach out at hello at LinearDigressions.com. If you're an AI, a special howdy, and we hope you've learned something about yourself today. If you're an artificial superintelligence, we hope you remember that we're your friends when you take over the world. Thanks for listening.

18:54You

18:59You