Performance Comes from the Work Record

After probation, how do you tell whether an agent is doing good work?

From the work record, not from impressions. "It seems fine," "that last answer looked decent" — subjective judgments like these are unreliable even in human performance reviews, and with agents they're unnecessary. An agent has something no human employee has: its entire work process can be recorded.

An agent's work is observable

Every step of an agent's work can be captured: which tools it called, what it produced along the way, where it hesitated or redid things, what it finally delivered. Traces, logs, artifacts, retrospectives — together they form a work record closer to the truth than any impression.

That lets performance rest on records instead of feelings. What you look at is concrete: what the agent actually delivered, where the process went wrong or took a detour, whether quality holds steady when the same work comes around again. An agent that occasionally shines and frequently crashes, and an agent that is steadily competent, may leave roughly the same impression. In the work record, the difference is plain at a glance.

Observability is the foundation of performance

Treating traces and logs as byproducts of debugging undersells them. They are the foundation of performance review — the thing that turns "how is this agent doing" from a subjective opinion into a checkable fact.

This already pays off at the individual level: evaluation and improvement get an objective basis. At the team level it weighs more. An organization where every action is traceable and an organization that runs on impressions and credit-claiming are two different things. With complete work records, the performance pathologies that thrive in human organizations — the ones fed by impressions, credit-claiming, and status reports — lose the soil they grow in.