Agent Evaluation

Production teams measure agents across four layers. Every layer catches different failure classes — benchmarks catch regressions on known tasks, traces catch silent drift, telemetry catches economic failure modes, human review catches what nobody wrote a rubric for.

No single benchmark is sufficient. Production teams run 3-4 benchmarks across different axes and reject any model that regresses on more than one.

Four common failure modes break the leaderboard-to-production correlation. Benchmark contamination leaks fixes into training data. Benchmark saturation makes the test stop discriminating. Distribution gap means the test set looks nothing like real users. Goodhart's law makes systems optimized for a benchmark regress on what the benchmark doesn't measure. The fix isn't fewer benchmarks — it's a portfolio diverse enough to resist overfitting, plus production telemetry as a sanity check.

Rubric-based LLM scoring, pioneered by Anthropic and adopted across labs, lets teams grade trajectories at scale without hand-annotating every turn. Anthropic's public guidance recommends pairing LLM-as-judge with spot-checked human review, especially on safety-sensitive categories. Judge bias toward verbose or confident-sounding answers is real — rubrics must explicitly correct for it. Production teams typically allocate 5-10% of trace volume to human review, weighted toward edge cases the judge flagged uncertainly.