AIhub coffee corner: Agentic AI

The AIhub coffee corner captures the musings of AI experts over a short conversation. This month we tackle the topic of agentic AI. Joining the conversation this time are: Sanmay Das (Virginia Tech), Tom Dietterich (Oregon State University), Sabine Hauert (University of Bristol), Sarit Kraus (Bar-Ilan University), and Michael Littman (Brown University).

Sabine Hauert: Today’s topic is agentic AI. What is it? Why is it taking off? Sanmay, perhaps you could kick off with what you noticed at AAMAS [the Autonomous Agents and Multiagent Systems conference]?

Sanmay Das: It was very interesting because obviously there’s suddenly been an enormous interest in what an agent is and in the development of agentic AI. People in the AAMAS community have been thinking about what an agent is for at least three decades. Well, longer actually, but the community itself dates back about three decades in the form of these conferences. One of the very interesting questions was about why everybody is rediscovering the wheel and rewriting these papers about what it means to be an agent, and how we should think about these agents. The way in which AI has progressed, in the sense that large language models (LLMs) are now the dominant paradigm, is almost entirely different from the way in which people have thought about agents in the AAMAS community. Obviously, there’s been a lot of machine learning and reinforcement learning work, but there’s this historical tradition of thinking about reasoning and logic where you can actually have explicit world models. Even when you’re doing game theory, or MDPs, or their variants, you have an explicit world model that allows you to specify the notion of how to encode agency. Whereas I think that’s part of the disconnect now – everything is a little bit black boxy and statistical. How do you then think about what it means to be an agent? I think in terms of the underlying notion of what it means to be an agent, there’s a lot that can be learnt from what’s been done in the agents community and in philosophy.

I also think that there are some interesting ties to thinking about emergent behaviors, and multi-agent simulation. But it’s a little bit of a Wild West out there and there are all of these papers saying we need to first define what an agent is, which is definitely rediscovering the wheel. So, at AAMAS, there was a lot of discussion of stuff like that, but also questions about what this means in this particular era, because now we suddenly have these really powerful creatures that I think nobody in the AAMAS community saw coming. Fundamentally we need to adapt what we’ve been doing in the community to take into account that these are different from how we thought intelligent agents would emerge into this more general space where they can play. We need to work out how we adapt the kinds of things that we’ve learned about negotiation, agent interaction, and agent intention, to this world. Rada Mihalcea gave a really interesting keynote talk thinking about the natural language processing (NLP) side of things and the questions there.

Sabine: Do you feel like it was a new community joining the AAMAS community, or the AAMAS community that was converting?

Sanmay: Well, there were people who were coming to AAMAS and seeing that the community has been working on this for a long time. So learning something from that was definitely the vibe that I got. But my guess is, if you go to ICML or NeurIPS, that’s very much not the vibe.

Sarit Kraus: I think they’re wasting some time. I mean, forget the “what is an agent?”, but there have been many works from the agent community for many years about coordination, collaboration, etc. I heard about one recent paper where they reinvented Contract Nets. Contract Nets were introduced in 1980, and now there is a paper about it. OK, it’s LLMs that are transferring tasks from one another and signing contracts, but if they just read the past papers, it would save their time and then they could move to more interesting research questions. Currently, they say with LLM agents that you need to divide the task into sub agents. My PhD was about building a Diplomacy player, and in my design of the player there were agents that each played a different part of a Diplomacy play – one was a strategic agent, one was a Foreign Minister, etc. And now they are talking about it again.

Michael Littman: I totally agree with Sanmay and Sarit. The way I think about it is this: this notion of “let’s build agents now that we have LLMs” to me feels a little bit like we have a new programming language like Rust++, or whatever, and we can use it to write programs that we were struggling with before. It’s true that new programming languages can make some things easier, which is great, and LLMs give us a new, powerful way to create AI systems, and that’s also great. But it’s not clear that they solve the challenges that the agents community have been grappling with for so long. So, here’s a concrete example from an article that I read yesterday. Claudius is a version of Claude and it was agentified to run a small online shop. They gave it the ability to communicate with people, post slack messages, order products, set prices on things, and people were actually doing economic exchanges with the system. At the end of the day, it was terrible. Somebody talked it into buying tungsten cubes and selling them in the store. It was just nonsense. The Anthropic people viewed the experiment as a win. They said “ohh yeah, there were definitely problems, but they’re totally fixable”. And the fixes, to me, sounded like all they’d have to do is solve the problems that the agents community has been trying to solve for the last couple of decades. That’s all, and then we’ve got it perfect. And it’s not clear to me at all that just making LLMs generically better, or smarter, or better reasoners suddenly makes all these kinds of agents questions trivial because I don’t think they are. I think they’re hard for a reason and I think you have to grapple with the hard questions to actually solve these problems. But it’s true that LLMs give us a new ability to create a system that can have a conversation. But then the system’s decision-making is just really, really bad. And so I thought that was super interesting. But we agents researchers still have jobs, that’s the good news from all this.

Sabine: My bread and butter is to design agents, in our case robots, that work together to arrive at desired emergent properties and collective behaviors. From this swarm perspective, I feel that over the past 20 years we have learned a lot of the mechanisms by which you reach consensus, the mechanisms by which you automatically design agent behaviours using machine learning to enable groups to achieve a desired collective task. We know how to make agent behaviours understandable, all that good stuff you want in an engineered system. But up until now, we’ve been profoundly lacking the individual agents’ ability to interact with the world in a way that gives you richness. So in my mind, there’s a really nice interface where the agents are more capable, so they can now do those local interactions that make them useful. But we have this whole overarching way to systematically engineer collectives that I think might make the best of both worlds. I don’t know at what point that interface happens. I guess it comes partly from every community going a little bit towards the other side. So from the swarm side, we’re trying visual language models (VLMs), we’re trying to have our robots understand using LLMs their local world to communicate with humans and with each other and get a collective awareness at a very local level of what’s happening. And then we use our swarm paradigms to be able to engineer what they do as a collective using our past research expertise. I imagine for those who are just entering this discipline they need to start from the LLMs and go up. I think it’s part of the process.

Tom Dietterich: I think a lot of it just doesn’t have anything to do with agents at all, you’re writing computer programs. People found that if you try to use a single LLM to do the whole thing, the context gets all messed up and the LLM starts having trouble interpreting it. In fact, these LLMs have a relatively small short-term memory that they can effectively use before they start getting interference among the different things in the buffer. So the engineers break the system into multiple LLM calls and chain them together, and it’s not an agent, it’s just a computer program. I don’t know how many of you have seen this system called DSPy (written by Omar Khattab)? It takes an explicit sort of software engineering perspective on things. Basically, you write a type signature for each LLM module that says “here’s what it’s going to take as input, here’s what it’s going to produce as output”, you build your system, and then DSPy automatically tunes all the prompts as a sort of compiler phase to get the system to do the right thing. I want to question whether building systems with LLMs as a software engineering exercise will branch off from the building of multi-agent systems. Because virtually all the “agentic systems” are not agents in the sense that we would call them that. They don’t have autonomy any more than a regular computer program does.

Sabine: I wonder about the anthropomorphization of this, because now that you have different agents, they’re all doing a task or a job, and all of a sudden you get articles talking about how you can replace a whole team by a set of agents. So we’re no longer replacing individual jobs, we’re now replacing teams and I wonder if this terminology also doesn’t help.

Sanmay: To be clear, this idea has existed at least since the early 90s, when there were these “soft bots” that were basically running Unix commands and they were figuring out what to do themselves. It’s really no different. What people mean when they’re talking about agents is giving a piece of code the opportunity to run its own stuff and to be able to do that in service of some kind of a goal.

I think about this in terms of economic agents, because that’s what I grew up (AKA, did my PhD) thinking about. And, do I want an agent? I could think about writing an agent that manages my (non-existent) stock portfolio. If I had enough money to have a stock portfolio, I might think about writing an agent that manages that portfolio, and that’s a reasonable notion of having autonomy, right? It has some goal, which I set, and then it goes about making decisions. If you think about the sensor-actuator framework, its actuator is that it can make trades and it can take money from my bank account in order to do so. So I think that there’s something in getting back to the basic question of “how does this agent act in the world?” and then what are the percepts that it is receiving?

I completely agree with what you were saying earlier about this question of whether the LLMs enable interactions to happen in different ways. If you look at pre-LLMs, with these agents that were doing pricing, there’s this hilarious story of how some old biology textbook ended up costing $17 million on Amazon because there were these two bots that were doing the pricing of those books at two different used book stores. One of them was a slightly higher-rated store than the other, so it would take whatever price that the lower-rated store had and push it up by 10%. Then the lower-rated store was an undercutter and it would take the current highest price and go to 99% of that price. But this just led to this spiral where suddenly that book cost $17 million. This is exactly the kind of thing that’s going to happen in this world. But the thing that I’m actually somewhat worried about, and anthropomorphising, is how these agents are going to decide on their goals.There’s an opportunity for really bad errors to come out of programming that wouldn’t be as harmful in a more constrained situation.

Tom: In the reinforcement learning literature, of course, there’s all this discussion about reward hacking and so on, but now we imagine two agents interacting with each other and hacking each other’s rewards effectively, so the whole dynamics blows up – people are just not prepared.

Sabine: The breakdown of the problem that Tom mentioned, I think there’s perhaps a real benefit to having these agents that are narrower and that as a result are perhaps more verifiable at the individual level, they maybe have clearer goals, they might be more green because we might be able to constrain what area they operate with. And then in the robotics world, we’ve been looking at collaborative awareness where narrow agents that are task-specific are aware of other agents and collectively they have some awareness of what they’re meant to be doing overall. And it’s quite anti-AGI in the sense that you have lots of narrow agents again. So part of me is wondering, are we going back to heterogeneous task-specific agents and the AGI is collective, perhaps? And so this new wave, maybe it’s anti-AGI – that would be interesting!

Tom: Well, it’s almost the only way we can hope to prove the correctness of the system, to have each component narrow enough that we can actually reason about it. That’s an interesting paradox that I was missing from Stuart Russell’s “What if we succeed?” chapter in his book, which is what if we succeed in building a broad-spectrum agent, how are we going to test it?

It does seem like it would be great to have some people from the agents community speak at the machine learning conferences and try to do some diplomatic outreach. Or maybe run some workshops at those conferences.

Sarit: I was always interested in human-agent interaction and the fact that LLMs have solved the language issue for me, I’m very excited. But the other problem that has been mentioned is still here – you need to integrate strategies and decision-making. So my model is you have LLM agents that have tools that are all sorts of algorithms that we developed and implemented and there should be several of them. But the fact that somebody solved our natural language interaction, I think this is really, really great and good for the agents community as well for the computer science community generally.

Sabine: And good for the humans. It’s a good point, the humans are agents as well in those systems.

AIhub is dedicated to free high-quality information about AI.

AIhub is dedicated to free high-quality information about AI.