Testing AI agents by asking them things at random turns out to be a poor strategy.
Researchers have introduced Monte Carlo Query Synthesis (MCQS), a method for probing black-box AI systems — foundation-model agents included — to learn what they can and cannot do. Rather than sending arbitrary inputs, MCQS uses Monte Carlo tree search to select queries that produce the most informative responses. The technique frames capability discovery as an active learning problem, targeting the gap between the most pessimistic and most optimistic explanations that fit what the system has already shown. Each query the system runs narrows that gap, so the method converges on an accurate capability model faster than random or baseline strategies. The researchers report soundness, completeness, and convergence guarantees under standard assumptions.
This matters because "deploy it and see what happens" is the de facto evaluation strategy for many AI systems today. As foundation-model agents take on multi-step tasks — booking travel, writing code, managing files — operators need a principled way to know the boundaries before something crosses them. MCQS offers a systematic alternative: a capability map built from evidence, not guesswork or cherry-picked demos.
The broader field has no shortage of benchmarks, but benchmarks measure fixed tasks. MCQS is trying to characterize open-ended behavior, which is a harder and arguably more useful problem — especially as regulators in the EU and elsewhere start asking vendors to prove their AI systems do what they claim.