zeroasterisk/agent-evals

Frames agent evaluations, benchmarks and leaderboards as held-constant ablations across three factors: Model, Harness, and Task Adaptation (none, pre-baked skill, or self-customization where the agent builds its own tools first). Use when deciding which eval cells to run next, designing a benchmark matrix, writing an eval report, or laying out a leaderboard UI so each product category gets a clear story without overwhelming viewers. Don't use for writing a single grader or task (build those first), or for generic statistics questions unrelated to agent evals.