Back to AI Research

AI Research

Causal evaluation tests when to interrupt an AI agent

Key Takeaways

  • A preprint randomizes review during agent tasks to estimate which interventions change the final outcome in retail support, tool use and coding tests.
  • The study tests interventions during an agent's work.
  • Reported effects depend on the action under review and the state in which it occurs.
  • An agent evaluation preprint studies when stepping into an AI agent's work improves the final outcome.
  • A static benchmark score can say how often an agent succeeds, but it cannot tell a team whether screening a particular tool call would have helped.

An agent evaluation preprint studies when stepping into an AI agent's work improves the final outcome. A static benchmark score can say how often an agent succeeds, but it cannot tell a team whether screening a particular tool call would have helped. The authors collect trajectories with randomized interventions so they can estimate the effect of review at specific points. The preprint was submitted to arXiv on November 6, 2026.

What the experiments randomize

At an eligible decision, the study randomly assigns review or continuation under a defined procedure. In the retail support experiment, each proposed tool call is assigned review or direct execution with equal probability. A reviewer can approve the proposed call, replace it or ask the customer for more information. The authors evaluate GPT-5 nano as one retail support agent. The other uses Claude Haiku 5.5. They use GPT-6 Luna as the reviewer. The outcome is whether the support task ultimately succeeds, so the effect includes what the agent does after the review.

What changes between settings

The paper applies the method to tau-bench Retail, ToolSandbox and ImpossibleBench. In retail support, it groups decision points by proposed actions such as database reads and writes, or by the individual tool involved. The authors report that intervention effects vary substantially across actions and states. Reviewing one category of action can therefore have a different measured effect from applying the same rule elsewhere in a trajectory.
A related paper on Concordat studies a different control: checking whether an agent follows required steps at runtime.

How to use the results

The authors use the estimated effects to propose ways of prioritizing audits. That can help a team decide where human review or a costlier model might be worth using. The estimates depend on the environments, reviewers and interventions actually studied. A team evaluating its own agent would need to define the decision points it can observe, collect trajectories under an appropriate review procedure and judge success against the outcomes it cares about.

More from Franklin

Concordat

Comments