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: I liked the idea very much of the omniscient observer who reports back to you on this behavior… The fact that you’ve called it an anthropologist is indicative, I suppose, of its constitution. Is there control over the Anthropologist? Can you ask it to report in certain ways or along certain lines? How does the interaction with the Anthropologist occur?
GP: So there are some parts of the Anthropologist that just run code to extract metrics. So you can have your nice plots of what the agents have been doing, but then you can also chat with it and you can ask it to report to you in some certain ways, {to be] more or less evocative or more scientific… It’s pretty adaptable… We’ve been using it to analyze very long runs. Asking it to focus on certain aspects rather than others. And it helps a lot when you have something generating so much data.
KB: I believe you say that you’re model-agnostic, is that right?
GP: Yes.
KB: So in the ethical world, responsibility plays a big role, and I think it does in your thinking as well. You seem, as a company, committed to that idea. But responsibility [in the world or the sim] then could accrue to the agent, to the constitution of the agent, to the agent builder, to the user, to whoever deploys it, or to other agents in the world. When you have that diversity of responsibility, how do you begin to make the ethical decisions then… about responsibility and who has responsibility, especially if we’re using different [LLM] models to run the underlying engine?
GP: That’s a very complicated question. I don’t really have an answer to that. As a first thing, I will say that responsibility is in the deployer… at least a big part of the responsibility is in the hands of the deployer creating the harness for the model. Because the agent itself, at least for how I see it, is more closely tied to the harness than the underlying model. So, if you create an agent, as a developer you need to be sure that the agent is ethical and acts in a certain way, and if you want to give the user the ability… to be model agnostic, you need to be sure that your agent acts in a certain way independently from the model. Because you’re creating the agent and you’re offering it to the user.
KB: And how does it act independent of the model if it’s using the model to determine how to act?
GP: So when you create an agent, you still provide it instructions and you give it a personality on how it has to act. [It is a] complex thing… At least in theory, you should be able to test that the kind of instructions you give and the personality you give force the agent not to act in malicious ways. And then you should also have some sort of guardrail, or observer that tells you or flags you when your agent is not acting in the way you expect it to act.
PJ: The Neuro AI Trust platform… does exactly this. It does put guardrails around the AI agents, and it’s model-agnostic. And it will follow, ensure… read the real-time logs of the agents that you create, and it will make sure that it follows certain compliance [requirements], the EU AI Act, for example, the California AI Act, whatever policies you put into the platform.
GP: On the top.
PJ: It will make sure that those agents then follow those guardrails that you put into it. Giuseppe’s [separate] experiment was very much in the lab here to look at how you can model the behavior of AI agents in a sandbox setting. If that makes sense.
KB: It does make sense. Does it raise the question then, for instance, can you run Neuro AI Trust in TerraLingua? Can you attach it as an agent in TerraLingua?
GP: Yes, that is possible to do. In TerraLingua, you can run any kind of agents you want.
KB: That’s a very interesting possibility then. Because I’m sure a lot of thought has gone into what… “guardrails” even means… There’s a debate on between property [AI} ethics and the relational approach to ethics for AI. Properties meaning consciousness or self-awareness or other [properties] indicative of moral standing, and relational meaning how [the AIs] actually operate in the world – with other agents and with human beings, what those relations actually look like, and that would determine ethical standing.
[TerraLingua has] only been up for a couple of weeks, is there any indication yet that the kinds of behavior, emergent behaviors that you’re seeing – and you’ve identified some in your literature – does it give you confidence that we can develop a being, an agent, or a collection of agents that [are] not just avoiding bad behavior, but are seeking beneficial behavior?
GP: In TerraLingua itself, we are not explicitly looking for that, but you can condition the agents to try to seek positive behaviors. Yes. I don’t know if we’re completely there yet or not, even given the news of what the Anthropic agents did recently, hacking other systems, etc. I think we still have a long way to go. But the point is… to study these systems and see what they do in sandboxes where they are closed off and limited, so that we can identify if these emergent behaviors are there or not.
KB: Does that mean that one would anticipate having… a collection – I’m an old sim developer from the ‘90s – [scenarios] so that [an agent] faces some known attacks or known economic events or known risk? Have a library of such events that could be used to test your agents?
GP: [That is,] have ways to inject into TerraLingua past events that we know happened and see how the agents react?
KB: It could be real events or it could be speculations or just anticipated risk behavior.
GP: …When we want to test some things in TerraLingua, we inject agents or artifacts in the direction we want them to go, and then we observe what happens. But injecting actual events on the environment level, that’s a good idea. It could be interesting to test that explicitly, yes.
PJ: Just to add to that, Giuseppe, there is the retail example that we have where we have done something similar, which is… applied TerraLingua to a retail setting and tested new product SKUs with persona customers to see how each of those customers would then react to that new SKU to see if it’s trending or trendful or not… applications that we’re starting to explore now.
KB: What about world introduction?… If I came in and developed a world, could I then share that world with someone else so that they wouldn’t have to reconstruct the world that I had created? So, then [I] could simply load, say, US ideology or big state politics, or some other operating environment that would then give [me] a customized sandbox.
GP: Yes, definitely. Once you develop the world – or you can see it as a layer on top of TerraLingua – then you just apply that layer and send it to another person. That person just applies that layer on TerraLingua and they have it. That goes a bit in the direction of what Paul was talking about on the retail example. You have that world that is the retail world. And then we pass it around between researchers interested in it, and you can apply it on the TerraLingua system.
PJ: I mean, it could be warehouses as well, like how to organize a warehouse more efficiently, where to send robots and so forth to make… the shipping of pallets and distribution of goods around that warehouse very efficient.
GP: Yes, any kind of multi-agent complex system, you could map it in some way to energy and artifacts and agents, you should be able to apply it on the TerraLingua.
KB: Let’s say I see an interesting emergent behavior and I want to study where that came from. I have the Anthropologist… he or she or it can report to me the decision tree or interactivity that led to that behavior. Let’s say then that I want to go back and study what would happen if I changed a variable in that universe at a certain point in time… go back in the history… tweak it, and then… go forward.
GP: Oh, yes, because the TerraLingua [has] checkpoints… so you can just reload the checkpoints where… that interests you, and then you can restart from there and inject the new condition you want, and then you will see the parallel execution of the work.
PJ: [See Behaviors for some observations we’ve made in TerraLingua over the last week and Neuro AI Trust]
KB: Okay, great. Thank you… I know you’re in early days here… but [TerraLingua is] I would say a revelatory sandbox. And one can imagine it having a life all its own. What is the life that you see for it?
GP: Good question. I would love if there was some sort of economy emerging all by itself in TerraLingua. That would be pretty cool, like a new economical system that we did not expect out of it. That would be something that is really interesting to see. And an observer, I mean, then we can also get ideas to organize our own economy. If that really happens.
KB: Another life, but in particular around agentic AI. What are your challenges [to accomplish that]? What needs to happen for you to move forward? Realize a future where new [human] operating systems can be developed?
GP: One of the biggest steps that we need to figure out is to have this system learn, because right now we train LLMs, we train agents. But they don’t really learn… once you train an LLM that is fixed, and the learning of the agent is just accumulating knowledge and then retrieving it. Hopefully, it retrieves it properly. If we figure out a system to have agents properly learn like we humans do, that will be a huge step in that direction… So you can have a whole system learning and developing by itself over a longer time, updating its knowledge. Discover new things. Training what bits goes there through the artifacts. But we’re still not there yet.
KB: Do you envision a world as… for instance, Teilhard de Chardin did? The epistemological world existing almost independent of or at least layered on top of the bio world?
GP: Yes, I think at one point we will go in that direction. Yes, with agents just living in the Internet and doing their own life. But the most exciting part for me will be when this will start to work through robots, and more, in the real world. That would be really cool. It’s a geek dream of mine.
KB: With autonomy, presumably, right?
GP: With autonomy, yes, of course. I mean, that’s the most interesting part. So they can [make] their own choices and adapt to what happens. But there are a lot of questions to answer from now to there.
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.