Demystifying evals for AI agents

Published 9 January 2026 by Mikaela Grace, Jeremy Hadfield, Rodrigo Olivares and Jiri De Jonghe, this Anthropic engineering guide begins by defining terms precisely, which is most of its value: an evaluation is a test that gives an AI an input and applies grading logic to its output, assembled from tasks, trials, graders, transcripts or traces, outcomes, an evaluation harness and an agent harness. It splits graders three ways. Code-based graders (string matching, binary tests, static analysis, outcome and tool-call verification) are fast, cheap, reproducible and brittle to valid variation. Model-based graders (rubric scoring, natural-language assertions, pairwise comparison, multi-judge consensus) are flexible and capture nuance but are non-deterministic, expensive and need calibration. Human graders are the gold standard and are what calibrates the other two. It then separates capability evals, which start at low pass rates and show where an agent is weak, from regression evals, which sit near 100 percent and catch backsliding, noting that a saturated capability eval graduates into the regression suite. Agent-type sections give distinct advice for coding agents (deterministic graders, SWE-bench Verified, Terminal-Bench), conversational agents (user simulation, tau-Bench and tau2-Bench), research agents (groundedness, coverage and source-quality checks, BrowseComp) and computer-use agents (sandboxed environments, WebArena, OSWorld). Non-determinism is handled with pass@k versus pass^k, identical at k=1 and telling opposite stories by k=10. An eight-step roadmap runs from 20-50 tasks drawn from real failures through harness isolation, grader design, transcript review and long-term ownership, and an appendix reviews Harbor, Braintrust, LangSmith, Langfuse and Arize.

Topic

Agentic

Level

Intermediate

Format

Guide

Cost

Free

Duration

~25 min read

Provider

Anthropic

Hands-on

No

Certificate

None

• ✓The precise anatomy of an eval: tasks, trials, graders, transcripts, outcomes and harnesses
• ✓When code-based, model-based and human graders each earn their cost
• ✓Separate capability evals from regression evals and graduate one into the other
• ✓Read pass@k against pass^k to expose an agent's real reliability
• ✓Apply distinct strategies for coding, conversational, research and computer-use agents
• ✓Follow an eight-step roadmap from 20-50 real failure cases to a maintained suite
• ✓Choose an eval framework among Harbor, Braintrust, LangSmith, Langfuse and Arize

• •Written by Anthropic engineers who run these evals internally, not a vendor pitch for a product
• •Names concrete benchmarks per agent type: SWE-bench Verified, Terminal-Bench, tau-Bench, tau2-Bench, BrowseComp, WebArena and OSWorld
• •The pass@k versus pass^k section is the clearest short treatment of agent non-determinism available
• •Insists on grading the outcome in the environment, not the final message, which is the defect most agent evals ship with
• •Step 0 is deliberately small, 20-50 tasks from real failures, so the roadmap is startable this week

This page was written from 2 sources, 1 on domains other than anthropic.com.

1. 1.anthropic.com — demystifying evals for ai agentsvendor
2. 2.ai-eval.org — anthropic demystifying evals for ai agents