How to Build an Agent Eval

Your agent gives you a research brief with footnotes and a sources section, and every claim has a little number beside it. Then you open one of those links and the page is dead. So you build an eval. Right instinct. And on its first draft, that checker will pass almost everything you put in front of it, the number will go up, and nothing will have got better. There is a smoke alarm on your kitchen ceiling that has never once gone off. You cannot tell whether that is because nothing has ever caught fire, or because the battery came out two years ago. Both look exactly like a quiet ceiling. That is the whole problem with evaluating agents: you are building an instrument, and an instrument that can never fail is not measuring anything. This video builds one end to end, in three parts. First the instrument: the tasks, the graders, the rubric. Then we watch it lie, because it will. Then we test it, which is the part almost nobody does. Covered: The vocabulary people use loosely: task, trial, outcome, transcript, grader, rubric, and the TWO different things both called a harness Why 20 to 50 cases drawn from real failures is a real answer, and not a concession The one-sentence test for a good task: would two domain experts, working separately, reach the same verdict? Grade the outcome, not the path, and why partial credit matters on multi-part tasks The grader ladder: code where the check is deterministic, an audited model where it is not, a person for the high stakes, and the cost logic that decides where each one lands The default failure mode of every rubric, and six things that fix it Why the thing that did the work cannot grade it, and what a fresh context window buys you Self-preference and judge instability, measured, and the calibration number that tells you when to trust a model judge: strongest at the extremes, weakest in the middle Harness configuration, and the compaction bug that silently scores a perfect run zero CORE-Bench: 42% to 95% with the model unchanged, because the grading and the task specs were fixed What an eval actually costs, and the three ways people pay more than they need to By the end there is an eleven-line card on screen. If somebody asks you in an interview how you would evaluate an agent, that card is your answer. Every claim in this video is sourced from Anthropic's Engineering blog, Alignment and Research posts, or the `anthropics/claude-cookbooks` repository, and each source is dated on screen rather than in the narration. This is a best-effort snapshot of the latest published material as of 8 September 2026. Guidance and APIs change; check the cookbooks and the docs before relying on any specific figure or API detail. Demystifying evals for AI agents (Anthropic Engineering, 9 Jan 2026) Bloom: Automated Behavioral Evaluations (Anthropic Alignment, 19 Dec 2025) Petri: an open-source auditing tool (Anthropic Research, 6 Oct 2025) How we built our multi-agent research system (Anthropic Engineering, 13 Jun 2025) https://www.anthropic.com/engineering... Building evals (cookbook, `anthropics/claude-cookbooks`) Verify with an outcome grader (Claude Managed Agents cookbook) Reproducing Claude's agentic search benchmark scores (cookbook) One thing is flagged on screen as practical advice rather than a citation: the effect of the ORDER you show two answers to a judge in. The mitigation costs nothing either way, so do it. 00:00 An alarm that never goes off 02:12 Task, trial, outcome, transcript, grader, rubric, harness 04:36 What goes in, and how many cases you need 06:42 Is it a good task? The two-expert test 08:31 Both directions: should, and should not 09:59 Grade the outcome, not the path 12:04 The grader ladder, and the cost logic 16:00 The vague rubric, and the grader that approves everything 18:16 Six rules for a rubric that can come back false 20:51 Those rules on one real run 23:14 Whatever did the work does not grade it 25:07 Your grader is a model too 27:26 The harness, and the question compaction loses 29:40 Four more settings, and three faults that are not yours 31:52 Sometimes the low score is your task 33:57 Testing the judge against people 36:31 Swiss cheese, and what time does to an eval 38:21 What an eval actually costs 40:43 The whole card, and the test button 43:01 Where to find more