My Agent Refused 96 Times. That Was the Right Output.

I wrote about a planning agent that was judged by the superficial metric of approval rate. This led to a failure as the agent refused almost every hard task, resulting in 96 out of 97 strict goals escalating. Initially, this seemed like failure, but it actually revealed a critical insight: a confident plan with a hidden flaw is often more dangerous than a clear refusal with a precise question.

The real lesson from PlannerCritic is that the safest output of an agent is sometimes not to produce a plan at all, but to stop and point out the blocker, forcing a human decision before the state changes.

The metric that initially raised concerns was that 96 of 97 strict goals escalated in v0.2.2, compared to 96 of 97 in v0.2.1. At first glance, this looked terrible, as if the system was failing almost every time a hard goal was presented. However, the key to understanding this lies in asking a better question: if a large share of those 96 escalations would have been worse than the refusals, which is a more dangerous output for a planning system to produce?

The most dangerous output is a plausible plan, often containing hidden caveats that hide the agent's uncertainty. The goal of an agent planner is not to produce the maximum number of plans, but to produce one of three trustworthy terminal states: approved, escalated, or blocked. Refusal is not inherently useful, but when it shrinks uncertainty into a precise next decision, it becomes a powerful tool.

In PlannerCritic, the refusal path is made explicit through the Escalation contract, escalation enforcement, and the loop controller that decides whether to approve, revise, or escalate. This ensures that humans are only involved when necessary, and the refusal is not just a decorative feature but a real, trustworthy path that reduces uncertainty and guides the user to the next decision point.