Demystifying evals for AI agents
Anthropic's practical guide to agent evals: grader types, pass@k vs pass^k, and a start-small roadmap (20-50 tasks from real failures). Teams with evals adopt new models in days instead of weeks.
**The gist** Anthropic published a practical guide to evaluating AI agents. It defines core terms (**pass@k vs pass^k**, trials, transcripts, graders), compares **code-based, model-based, and human graders**, and walks through eval patterns for coding, conversational, research, and computer-use agents, citing benchmarks like **SWE-bench Verified** and **τ-Bench**.
**Why it matters** The roadmap is directly usable: start with **20-50 tasks** drawn from real failures, prefer deterministic graders, read transcripts to check your graders, and retire saturated evals into regression suites. Anthropic claims teams with evals adopt new models **in days instead of weeks** because regressions surface immediately.
**The gist** Anthropic published a practical guide to evaluating AI agents. It defines core terms (**pass@k vs pass^k**, trials, transcripts, graders), compares **code-based, model-based, and human graders**, and walks through eval patterns for coding, conversational, research, and computer-use agents, citing benchmarks like **SWE-bench Verified** and **τ-Bench**. **Why it matters** The roadmap is directly usable: start with **20-50 tasks** drawn from real failures, prefer deterministic graders, read transcripts to check your graders, and retire saturated evals into regression suites. Anthropic claims teams with evals adopt new models **in days instead of weeks** because regressions surface immediately. **Watch out** Eval bugs distort results — a grading fix moved **Opus 4.5 from 42% to 95%** on CORE-Bench, and a **0% pass rate** usually means a broken task, not a weak agent. LLM judges need human calibration, and evals miss creative agent behavior, so pair them with production monitoring.