AIhub espresso nook: Agentic AI

The AIhub espresso nook captures the musings of AI consultants over a brief dialog. This month we deal with the subject of agentic AI. Becoming a member of the dialog this time are: Sanmay Das (Virginia Tech), Tom Dietterich (Oregon State College), Sabine Hauert (College of Bristol), Sarit Kraus (Bar-Ilan College), and Michael Littman (Brown College).

Sabine Hauert: As we speak’s matter is agentic AI. What’s it? Why is it taking off? Sanmay, maybe you could possibly kick off with what you observed at AAMAS [the Autonomous Agents and Multiagent Systems conference]?

Sanmay Das: It was very fascinating as a result of clearly there’s instantly been an infinite curiosity in what an agent is and within the improvement of agentic AI. Folks within the AAMAS neighborhood have been interested by what an agent is for not less than three many years. Effectively, longer truly, however the neighborhood itself dates again about three many years within the type of these conferences. One of many very fascinating questions was about why everyone is rediscovering the wheel and rewriting these papers about what it means to be an agent, and the way we should always take into consideration these brokers. The best way during which AI has progressed, within the sense that giant language fashions (LLMs) are actually the dominant paradigm, is nearly solely completely different from the way in which during which individuals have thought of brokers within the AAMAS neighborhood. Clearly, there’s been lots of machine studying and reinforcement studying work, however there’s this historic custom of interested by reasoning and logic the place you may even have specific world fashions. Even if you’re doing sport concept, or MDPs, or their variants, you may have an specific world mannequin that permits you to specify the notion of learn how to encode company. Whereas I feel that’s a part of the disconnect now – every little thing is just a little bit black boxy and statistical. How do you then take into consideration what it means to be an agent? I feel by way of the underlying notion of what it means to be an agent, there’s rather a lot that may be learnt from what’s been carried out within the brokers neighborhood and in philosophy.

I additionally suppose that there are some fascinating ties to interested by emergent behaviors, and multi-agent simulation. Nevertheless it’s just a little little bit of a Wild West on the market and there are all of those papers saying we have to first outline what an agent is, which is certainly rediscovering the wheel. So, at AAMAS, there was lots of dialogue of stuff like that, but additionally questions on what this implies on this specific period, as a result of now we instantly have these actually highly effective creatures that I feel no one within the AAMAS neighborhood noticed coming. Basically we have to adapt what we’ve been doing locally to keep in mind that these are completely different from how we thought clever brokers would emerge into this extra normal area the place they’ll play. We have to work out how we adapt the sorts of issues that we’ve realized about negotiation, agent interplay, and agent intention, to this world. Rada Mihalcea gave a extremely fascinating keynote speak interested by the pure language processing (NLP) facet of issues and the questions there.

Sabine: Do you’re feeling prefer it was a brand new neighborhood becoming a member of the AAMAS neighborhood, or the AAMAS neighborhood that was changing?

Sanmay: Effectively, there have been individuals who had been coming to AAMAS and seeing that the neighborhood has been engaged on this for a very long time. So studying one thing from that was positively the vibe that I bought. However my guess is, in case you go to ICML or NeurIPS, that’s very a lot not the vibe.

Sarit Kraus: I feel they’re losing a while. I imply, neglect the “what’s an agent?”, however there have been many works from the agent neighborhood for a few years about coordination, collaboration, and so on. I heard about one latest paper the place they reinvented Contract Nets. Contract Nets had been launched in 1980, and now there’s a paper about it. OK, it’s LLMs which might be transferring duties from each other and signing contracts, but when they simply learn the previous papers, it will save their time after which they might transfer to extra fascinating analysis questions. Presently, they are saying with LLM brokers that you could divide the duty into sub brokers. My PhD was about constructing a Diplomacy participant, and in my design of the participant there have been brokers that every performed a unique a part of a Diplomacy play – one was a strategic agent, one was a International Minister, and so on. And now they’re speaking about it once more.

Michael Littman: I completely agree with Sanmay and Sarit. The best way I give it some thought is that this: this notion of “let’s construct brokers now that we’ve got LLMs” to me feels just a little bit like we’ve got a brand new programming language like Rust++, or no matter, and we will use it to write down packages that we had been fighting earlier than. It’s true that new programming languages could make some issues simpler, which is nice, and LLMs give us a brand new, highly effective solution to create AI techniques, and that’s additionally nice. Nevertheless it’s not clear that they resolve the challenges that the brokers neighborhood have been grappling with for therefore lengthy. So, right here’s a concrete instance from an article that I learn yesterday. Claudius is a model of Claude and it was agentified to run a small on-line store. They gave it the power to speak with individuals, submit slack messages, order merchandise, set costs on issues, and other people had been truly doing financial exchanges with the system. On the finish of the day, it was horrible. Any individual talked it into shopping for tungsten cubes and promoting them within the retailer. It was simply nonsense. The Anthropic individuals seen the experiment as a win. They mentioned “ohh yeah, there have been positively issues, however they’re completely fixable”. And the fixes, to me, seemed like all they’d should do is resolve the issues that the brokers neighborhood has been attempting to resolve for the final couple of many years. That’s all, after which we’ve bought it excellent. And it’s not clear to me in any respect that simply making LLMs generically higher, or smarter, or higher reasoners instantly makes all these sorts of brokers questions trivial as a result of I don’t suppose they’re. I feel they’re laborious for a motive and I feel it’s important to grapple with the laborious questions to truly resolve these issues. Nevertheless it’s true that LLMs give us a brand new potential to create a system that may have a dialog. However then the system’s decision-making is simply actually, actually dangerous. And so I assumed that was tremendous fascinating. However we brokers researchers nonetheless have jobs, that’s the excellent news from all this.

Sabine: My bread and butter is to design brokers, in our case robots, that work collectively to reach at desired emergent properties and collective behaviors. From this swarm perspective, I really feel that over the previous 20 years we’ve got realized lots of the mechanisms by which you attain consensus, the mechanisms by which you mechanically design agent behaviours utilizing machine studying to allow teams to attain a desired collective job. We all know learn how to make agent behaviours comprehensible, all that great things you need in an engineered system. However up till now, we’ve been profoundly missing the person brokers’ potential to work together with the world in a manner that provides you richness. So in my thoughts, there’s a very nice interface the place the brokers are extra succesful, to allow them to now do these native interactions that make them helpful. However we’ve got this complete overarching solution to systematically engineer collectives that I feel would possibly make the perfect of each worlds. I don’t know at what level that interface occurs. I assume it comes partly from each neighborhood going just a little bit in the direction of the opposite facet. So from the swarm facet, we’re attempting visible language fashions (VLMs), we’re attempting to have our robots perceive utilizing LLMs their native world to speak with people and with one another and get a collective consciousness at a really native stage of what’s taking place. After which we use our swarm paradigms to have the ability to engineer what they do as a collective utilizing our previous analysis experience. I think about for individuals who are simply getting into this self-discipline they should begin from the LLMs and go up. I feel it’s a part of the method.

Tom Dietterich: I feel lots of it simply doesn’t have something to do with brokers in any respect, you’re writing pc packages. Folks discovered that in case you attempt to use a single LLM to do the entire thing, the context will get all tousled and the LLM begins having hassle decoding it. Actually, these LLMs have a comparatively small short-term reminiscence that they’ll successfully use earlier than they begin getting interference among the many various things within the buffer. So the engineers break the system into a number of LLM calls and chain them collectively, and it’s not an agent, it’s simply a pc program. I don’t know what number of of you may have seen this method referred to as DSPy (written by Omar Khattab)? It takes an specific form of software program engineering perspective on issues. Principally, you write a kind signature for every LLM module that claims “right here’s what it’s going to take as enter, right here’s what it’s going to supply as output”, you construct your system, after which DSPy mechanically tunes all of the prompts as a form of compiler part to get the system to do the suitable factor. I wish to query whether or not constructing techniques with LLMs as a software program engineering train will department off from the constructing of multi-agent techniques. As a result of just about all of the “agentic techniques” usually are not brokers within the sense that we might name them that. They don’t have autonomy any greater than an everyday pc program does.

Sabine: I ponder in regards to the anthropomorphization of this, as a result of now that you’ve completely different brokers, they’re all doing a job or a job, and impulsively you get articles speaking about how one can exchange an entire staff by a set of brokers. So we’re not changing particular person jobs, we’re now changing groups and I ponder if this terminology additionally doesn’t assist.

Sanmay: To be clear, this concept has existed not less than because the early 90s, when there have been these “gentle bots” that had been principally operating Unix instructions they usually had been determining what to do themselves. It’s actually no completely different. What individuals imply once they’re speaking about brokers is giving a bit of code the chance to run its personal stuff and to have the ability to try this in service of some type of a purpose.

I take into consideration this by way of financial brokers, as a result of that’s what I grew up (AKA, did my PhD) interested by. And, do I need an agent? I might take into consideration writing an agent that manages my (non-existent) inventory portfolio. If I had sufficient cash to have a inventory portfolio, I’d take into consideration writing an agent that manages that portfolio, and that’s an inexpensive notion of getting autonomy, proper? It has some purpose, which I set, after which it goes about making choices. If you concentrate on the sensor-actuator framework, its actuator is that it might make trades and it might take cash from my checking account so as to take action. So I feel that there’s one thing in getting again to the essential query of “how does this agent act on this planet?” after which what are the percepts that it’s receiving?

I utterly agree with what you had been saying earlier about this query of whether or not the LLMs allow interactions to occur in several methods. Should you have a look at pre-LLMs, with these brokers that had been doing pricing, there’s this hilarious story of how some outdated biology textbook ended up costing $17 million on Amazon as a result of there have been these two bots that had been doing the pricing of these books at two completely different used e-book shops. Considered one of them was a barely higher-rated retailer than the opposite, so it will take no matter value that the lower-rated retailer had and push it up by 10%. Then the lower-rated retailer was an undercutter and it will take the present highest value and go to 99% of that value. However this simply led to this spiral the place instantly that e-book price $17 million. That is precisely the type of factor that’s going to occur on this world. However the factor that I’m truly considerably frightened about, and anthropomorphising, is how these brokers are going to resolve on their targets.There’s a possibility for actually dangerous errors to come back out of programming that wouldn’t be as dangerous in a extra constrained scenario.

Tom: Within the reinforcement studying literature, in fact, there’s all this dialogue about reward hacking and so forth, however now we think about two brokers interacting with one another and hacking one another’s rewards successfully, so the entire dynamics blows up – persons are simply not ready.

Sabine: The breakdown of the issue that Tom talked about, I feel there’s maybe an actual profit to having these brokers which might be narrower and that because of this are maybe extra verifiable on the particular person stage, they possibly have clearer targets, they is perhaps extra inexperienced as a result of we’d have the ability to constrain what space they function with. After which within the robotics world, we’ve been taking a look at collaborative consciousness the place slender brokers which might be task-specific are conscious of different brokers and collectively they’ve some consciousness of what they’re meant to be doing general. And it’s fairly anti-AGI within the sense that you’ve numerous slender brokers once more. So a part of me is questioning, are we going again to heterogeneous task-specific brokers and the AGI is collective, maybe? And so this new wave, possibly it’s anti-AGI – that may be fascinating!

Tom: Effectively, it’s nearly the one manner we will hope to show the correctness of the system, to have every part slender sufficient that we will truly motive about it. That’s an fascinating paradox that I used to be lacking from Stuart Russell’s “What if we succeed?” chapter in his e-book, which is what if we reach constructing a broad-spectrum agent, how are we going to check it?

It does seem to be it will be nice to have some individuals from the brokers neighborhood communicate on the machine studying conferences and attempt to do some diplomatic outreach. Or possibly run some workshops at these conferences.

Sarit: I used to be at all times fascinated by human-agent interplay and the truth that LLMs have solved the language concern for me, I’m very excited. However the different drawback that has been talked about remains to be right here – you could combine methods and decision-making. So my mannequin is you may have LLM brokers which have instruments which might be all types of algorithms that we developed and carried out and there ought to be a number of of them. However the truth that any individual solved our pure language interplay, I feel that is actually, actually nice and good for the brokers neighborhood as effectively for the pc science neighborhood typically.

Sabine: And good for the people. It’s an excellent level, the people are brokers as effectively in these techniques.