Ask any deployed agent a question to which it does not know the answer, and watch carefully what happens. In a meaningful fraction of cases, it will produce an answer anyway. The answer will be plausible, fluent, and confidently delivered. It will also, often enough, be wrong.

This behaviour is not, despite the popular framing, a bug introduced by training. It is closer to a structural feature of systems whose objective is to produce coherent text. Coherence and accuracy are correlated, but they are not the same thing, and when they diverge, coherence has historically won.

The work that has impressed me most this year takes the problem seriously at the level of architecture. Several teams have been building agents whose internal state explicitly tracks calibration: how confident the model believes it should be, how confident it actually is, and the gap between the two. When the gap exceeds a threshold, the agent is structurally prevented from emitting an unhedged claim. It must either gather more information or admit uncertainty.

The early systems are, by the standards of frontier deployments, slow and occasionally infuriating. They ask clarifying questions when a snappier system would have guessed. They refuse tasks that a more cavalier model would have attempted and gotten right four times out of five. The fifth time, however, is the one that matters, and the calibrated systems do not have a fifth time.

Whether this approach scales, and whether users will tolerate its texture once they grow accustomed to fluency, are open questions. What seems no longer open is the older claim that honesty is simply a matter of training harder on better data. The architecture has to want it.

What an Honest Agent Looks Like, and Why We Have So Few