KB Miller, publisher at AGI Ethics News, talks with Cognizant about its testing ground for agentic AI – just launched
Agentic AI joins artificial intelligence with digital programming that interacts with and operates on its environment. These agents are on the frontier of AI development, presenting unprecedented opportunities for leveraging human productivity along with unpredictable challenges of safety and risk. In its AI R&D lab, led by Babak Hodjat, this large company has set up – for its own and any outside researcher’s use – a tester’s sandbox. Called “the first living AI ecology,” TerraLingua tests agentic AIs before they are deployed commercially or released in public-facing systems. Watching and experimenting with the behavior of these living systems then becomes a formative part of the agent design process. We talked with Giuseppe Paolo, a driving force behind TerraLingua’s development, and Paul Jarratt, Head of Thought Leadership.
Giuseppe Paolo is a research scientist at Cognizant AI Lab and lead creator of TerraLingua. He studies agentic, multi-agent, and embodied AI and holds a PhD from Sorbonne University.
Paul Jarratt is Head of Thought Leadership for Cognizant AI Lab, driving content, communications, and visibility across Cognizant’s AI research portfolio.
This interview has been edited for clarity and readability.
KB Miller: So I think the place to start is just introduce Cognizant and TerraLingua to our readers. Cognizant is one of the leading firms in the world in agentic AI development. We should know about Cognizant and its place in this universe.
Giuseppe Paolo: Cognizant is trying to be the leader in agentic AI. So we, as the AI Lab, are trying to develop the technology, or perform the science, to get there. And TerraLingua is one of the big parts of [that] project… to have these multi-agent approaches, a system where we have multiple agents that can collaborate together… in our network and for the whole community in general.
Personally, I’m mainly working on TerraLingua and the way to develop it scientifically, and seeing how this could be… [realized] in the world.
KB: We hear every day now about [the advent of] artificial general intelligence [AGI]. As a leader in this field, do you feel that agentic AI is required for reaching AGI or indeed ASI.
GP: I would say so, because I’m a strong believer of embodiment for AI in general, so that it needs to live in its own environment and needs to be able to interact with the world it lives in, so to do that, you need agency, so agentic AI is a big component of that. I don’t really believe that just using one big model, one big LLM, will solve AGI, will get us there… We really need a way for the system to be able to interact with its world, to learn from it, to get some feedback from it. And because otherwise you’re just a passive observer, you cannot do anything.
KB: That’s the difference between telling and doing, right?
GP: Yes, exactly. I mean, with all the intelligence we know about, it’s really an active intelligence, so it does things in the world it lives in… We humans are [agentic intelligence], but even animals… they’re really agentic in the world they live in.
KB: So… individual agents are going to be necessary to create a true AGI… is that [theory] still evolving or do we know pretty much where to go, we just haven’t done it yet?
GP: No, I don’t think we have identified the main, the principal agents… We still have a lot to discover, many things to figure out. For now, I think. Many people are still pushing this… a big single agent approach that can get to AGI. But if… AGI will be a whole system of multiple parts, some of which we still don’t really know yet… the kind of agents we will need that really depends on the task you want to solve.
So probably AGI will not be fixed agents, but will be a system of adaptive agents that will create new ones depending on the task it has to solve and delete old ones.
KB: That makes sense. Go back to your thought about embodiment for a minute. Are you… talking about a cyber embodiment in a digital world, or are you talking about, perhaps, a physical embodiment so that robotics might be required to deliver true AGI.
“I actually think it’s even better to study ethical behaviors in sandboxes, so you don’t… break people or things in the real world… But in a sandbox, if you manage to constrain properly, to simulate other humans properly in the sandbox, I think that’s the way to go.”
GP: I think physical embodiment is not really necessary. It can help a lot because… you don’t have to build a simulator to get your data. You have everything there, and we’re much more familiar with it as well, because we live in a physical world. But I think a digital embodiment is already good enough, I will say. I mean, an agent living on the Internet has access to all our knowledge and can perform any sort of task. So it should be enough… to make agents smarter.
KB: And is that world yet in TerraLingua or is that a world model that has to be attached?
GP: So it’s not in TerraLingua. TerraLingua is a relatively simple world, because we wanted to study some pretty precise questions with it. [The digital or physical world is] the place where you will train your AGI. I think a world model would be a component of an AGI system, but this world model will need to be learned on the environment where the agent is embodied. I don’t know which would be the best embodiment world for such a system. This is why, actually, the real physical work could make things easier because you already have all the complexity you need.
But then you need to deal with hardware and robotics, and that’s a whole mess of things to make work.
KB: Yes, considering that [for] the human body we’ve identified something like 30 plus sensory systems. It would be a complex task just to collect the data that you need to have a human experience in any case.
GP: Yes, exactly. I’ve been away from the robotics world for awhile, but even just making the skin that is a super complex perception system is very hard as a robot. So… they’re starting to develop it. But having that is super hard, and all the other sensing systems as well. It’s not easy.
KB: In our world at AGI Ethics News, we’re deeply concerned with the ethical [AI] issues. Do you agree that operating in an ethical world could be done in a digital world? Or at least we could get suggestive behavior out of a simulation… for studying human ethical behavior. Do you agree with that? Do you see that as a possible direction or [that] one of the applications of the sandbox is studying ethical behavior?
GP: Yes, I actually think it’s even better to study ethical behaviors in sandboxes, so you don’t… break people or things in the real world… But in a sandbox, if you manage to constrain properly, to simulate other humans properly in the sandbox, I think that’s the way to go.
KB: Tell us a little about the TerraLingua environment, then. What is the world that you inherit when you come into TerraLingua?
GP: You mean as an agent? What is the world you inherit?
KB: As an agent and as a developer of agents.
GP: Okay, so as an agent it’s a world where your ancestors lived. They… spent their whole life in it. They created things. Sometimes they create epitaphs or tombstones for themselves or leave diaries behind. So you have a whole history. Or what has happened in the world that you can go around and reconstruct. As an agent, you find yourself in this continuously evolving world, very rich of history, of culture from the past generations which is pretty similar, I would say, to how a human born in the real world… finds itself with culture spun from generations.
“It’ll then monitor and in real time stop or prevent agent behavior that may be misbehaving, depending on certain policies and governance that you put around it.”
While as an agent developer, if you enter into running the TerraLingua simulation, you also find that [culture], then you can adapt your agent… make your agent more or less reactive to certain things, depending on what your goal is. And you can go around studying the history, the culture of the world… trying to understand why some agents did certain things, why some agents did other things, and you can decide how you want your agents to be.
KB: It seems like one of the challenges in model building has been to catch errors or bad behavior before it happens. In a world where the agents actually do… in addition to tell… does there appear to be promise that you can catch the bad behavior in advance of deployment? And how does your approach to guardrails work? [Does it] differ because they are agentic AI?
GP: Yes, so there is this promise… Try your agent in the world, see how it behaves depending on the setting you put it in, what the… other inputs you give it, etc. And then you can see how it behaves because, at least from my experience, you cannot know beforehand how an agent with a given harness will react to the world. So you need to actually test it. And the fact that you have sandboxes like this allows you to test it live before actually [deploying], and test your assumptions… what your agent will do.
[With] a system like TerraLingua, you can test your models, but you can also test your harnesses yourself in a setting where otherwise you need to simulate behaviors and actions.
Paul Jarratt: KB, just on your guardrails question, I think you also spoke to Amir [Banifatemi], our Chief Responsible AI Officer, potentially about our Neuro AI Trust product, which is more about how you can put guardrails around agents and register agents within an enterprise, to apply certain policies. (Neuro AI Trust) gives Cognizant clients real-time governance over every connected agent across any AI stack using an SDK (software development kit) to automatically register them, so companies can monitor, control, and trust agentic AI as it operates at enterprise scale.]
It’ll then monitor and in real time stop or prevent agent behavior that may be misbehaving, depending on certain policies and governance that you put around it. That’s obviously different from TerraLingua, but it’s something that we’re offering clients.
KB: Talk for a minute then about if TerraLingua links into that process. Does TerraLingua become an area where Cognizant can essentially give a developer… a kind of “seal of approval” or a measure of competence of an agent that can be used to determine whether the agent is releasable or not? Is there some measure of maturity, if you will, of an agent that gives confidence to the developer? Are they doing all that themselves? Or is Cognizant a partner in that work?
PJ: So for the Neuro AI Trust… this is more for semi-autonomous agents. You offload to a human-in-the-loop for certain tasks, right? Whereas I think Giuseppe’s experiment is more about fully autonomous agents, how do you monitor and measure their behavior in a sandbox environment?
GP: Yes. And then you can have a human check the results afterwards, depending on what you expect from your agent.
KB: Right. So there’s no human-in-the-loop in Terra Lingua.
GP: No.
KB: Right. But there could be, in theory, an ethical-agent-in-the-loop, right?
GP: Yes, definitely. You could have an agent observer that tells the agents in the simulation that they’re doing well, or they’re unethical, or they’re not acting as you expect.
We already have something similar that we call the AI Anthropologist. That is an agent observer that observes what the other agent did and then gives a report to the user. So it does not interact… with the simulation itself. But having something like that that interacts with the simulator and drives it, that’s definitely possible. Yes.
Don’t miss Part 2 coming next month! Be sure to check out an exclusive interview with AGI Ethics News publisher KB Miller and NYU’s Toni Sims here where they discuss AI welfare.
AGI Ethics News covers the ethical and societal implications of artificial general intelligence (AGI). We bring expert insights and thoughtful analysis to help readers understand the challenges and responsibilities of emerging AI technologies.