How Om Buddhdev Is Making AI Behavior Something Researchers Can Inspect

Two AI systems can reach the same result for very different reasons. Om Buddhdev, co-founder and CEO of Olam Labs, is building evaluation environments designed to preserve those differences rather than reduce them to a single outcome. His work focuses on making social and agentic behavior measurable by controlling the conditions participants face and retaining enough information to examine how an interaction developed. That approach gives researchers a way to study not only performance, but the circumstances and decisions that shaped it.

"A result can tell you whether something worked, but it cannot always tell you why," Buddhdev says. "If you want to understand the behavior of the model, you need enough information to reconstruct what happened before the outcome."

One of the most important design choices in Buddhdev's work is human-agent parity. Olam Labs places humans and frontier AI models into social games where they receive equivalent information and have access to equivalent actions. That structure gives researchers a more controlled basis for comparing how humans and models behave within the same environment.

"We give humans and agents equivalent information and available actions," Buddhdev says. "That gives us a consistent environment for studying how different participants behave."

The environments also preserve full traces of what happens during an interaction. Communications, actions, system state, incentives, uncertainty, and outcomes can all be retained for analysis. Buddhdev uses that information to study behaviors such as negotiation, collaboration, deception, memory, and adaptation with more context than a final score can provide. The objective is to make the path through the interaction visible enough that researchers can investigate what contributed to a result.

That emphasis is closely tied to what Buddhdev describes as causal legibility. In practice, it means retaining enough of the environment and interaction history to examine which decisions, conditions, or responses contributed to an outcome. A failed collaboration, for example, might come from a misunderstood incentive, a missed signal, or a choice that changes how another participant responds. Those causes are not interchangeable, even when they eventually produce the same result.

Buddhdev's environments are designed to keep those causes available for analysis. Olam Labs creates controlled but emergent settings in which communications, actions, state, incentives, uncertainty, and outcomes can all be inspected. The rules of the environment provide structure, while the interaction itself can still produce unexpected situations. That balance allows Buddhdev and his team to study complicated behavior without losing the information needed to understand how it developed.

"The interaction has to stay measurable even when the behavior becomes complicated," he says. "You want to preserve what was communicated, what actions were available, how the state changed, and what incentives were present. Otherwise, you may know what happened without being able to explain the conditions around it."

This methodology becomes especially useful when researchers want to compare different strategies under similar conditions. One participant may respond quickly to a change in incentives, while another may continue using an approach that no longer fits the situation. A human participant may interpret an uncertain exchange differently from an AI agent facing the same available information. Because the environment is structured around comparable conditions, those differences become more useful as evidence about behavior rather than simply anecdotes about performance.

The same design also gives researchers more precise questions to ask. They can examine whether a model's communication changes after another participant alters its strategy. They can study whether earlier information is still influencing a decision several steps later. They can also look at whether a breakdown in cooperation followed from the model's own actions or from a change elsewhere in the environment. The method is built around preserving enough context to support those distinctions.

Buddhdev sees that level of visibility as increasingly important as AI systems take on tasks that unfold across longer periods and involve more independent action. In those settings, a completed task may reveal very little about the quality of the behavior that produced it. A system could reach an acceptable result after a series of poor choices, or it could fail after making several reasonable decisions under difficult conditions. Understanding the difference requires more than an endpoint.

For Buddhdev, the value of these environments depends on whether the behavior inside them remains measurable. Social games and multi-agent scenarios may look dynamic on the surface, but their usefulness comes from the information they produce about how participants respond under changing conditions. The environment must allow meaningful interaction while preserving enough structure for researchers to analyze what happened.

This emphasis also shapes how he thinks about evaluation design more broadly. Buddhdev argues that researchers should instrument systems well enough to understand why outcomes occur and keep claims closely tied to the evidence available. In the context of Olam Labs, that means building environments where the behavior itself can be examined rather than inferred from a final score.

"Good evaluation should give you more than a ranking," Buddhdev says. "It should give you enough context to understand what the system actually did and what conditions influenced that behavior."

As AI agents become more capable of interacting with people and one another, Buddhdev believes the need for that context will grow. Social and agentic behavior is difficult to assess when the underlying interaction disappears after the task is complete. Preserving comparable conditions, full traces, and the surrounding state gives researchers a stronger basis for examining what happened.

That is the central purpose of the environments Olam Labs is building. The goal is to make complex AI behavior observable without stripping away the uncertainty and interaction that make it worth studying in the first place. By combining human-agent parity with detailed traces and causal legibility, he is building evaluation systems intended to give researchers more evidence about how AI behavior develops inside complex interactions.