The video is narrated in Japanese. This page is the English write-up: the same argument, key points, full transcript in translation, FAQ and primary sources.
The short answer
An AI agent should not be judged on its final reply alone. Score the state it left behind in the environment, the tool calls it made along the way, its safety, and its stability across repeated trials — separately. A practical start is 20-50 tasks drawn from real failures and business-critical work, with the success state and the forbidden actions written down explicitly.
Key points
- A task is the problem, a trial is one attempt at it, and a grader is the scoring logic.
- The transcript is the record of what the agent did; the outcome is the state finally left in the environment.
- Generative models vary run to run, so measure stability across several trials rather than trusting one success.
- Score outcome, process, safety and efficiency separately instead of collapsing them into a single number.
- Turn your characteristic failures into 20-50 tasks and keep growing the evaluation set.
Transcript
Read the full transcript
Anthropic's "Demystifying evals for AI agents", published on 9 January 2026, is a practitioner's write-up on how to evaluate agents that use tools, decide repeatedly, and change their environment.
For a single-turn answer you can simply compare input with output. An agent, though, uses tools across many turns and changes its behaviour based on intermediate results, so what you evaluate has to widen from the final sentence to the whole job.
A task is one problem; a trial is one attempt; a grader is the scoring logic; a transcript is the complete record including replies and tool calls; an outcome is the state left in the environment at the end. The evaluation harness handles running, recording, scoring and aggregating.
A travel agent saying "I booked it" is not proof of success. You verify the outcome from the outside: does the booking database actually hold the right flight on the right terms?
Running a task once tells you nothing about stability. Use several trials to see the success rate and the shape of the failures, so you do not mistake a lucky run for real capability.
Use separate graders for the final result, the required conditions, the forbidden operations, cost and tool calls, and hand-off to a human. A single combined score erases the difference between an agent that is fast but dangerous and one that is slow but safe.
The cause of a failure is in the transcript. Which information it read, which tool it chose, where it misread a condition, and what it overwrote — investigating that is what makes improvement and auditing possible.
The evaluation environment needs mock tickets, test records, anonymised data — somewhere the agent can genuinely change state. You are measuring whether the right change was left behind, not whether the prose was elegant.
A capable agent will sometimes find a correct answer your static rubric never anticipated. Do not rely on the rubric alone; combine it with human review and with what the result meant for the user.
You do not need a large platform to begin. Collect 20-50 tasks you genuinely cannot afford to fail, mix in typical, boundary and dangerous cases, and add incidents and user complaints as new tasks.
Sales, accounting, HR and legal each have different good outcomes and different forbidden actions. Evaluating AI is not scoring a model's IQ; it is the work of defining what a good ending to a job looks like for your company.
FAQ
What should you evaluate in an AI agent?
Score these separately: the final answer, the state left in the environment, the tool calls made along the way, safety, efficiency, hand-off to a human, and stability across repeated trials.
What is the difference between a transcript and an outcome?
The transcript is the full record of one trial including replies and tool calls; the outcome is the state finally left in the environment, such as in a database or a file.
How many evaluation tasks should you start with?
Anthropic describes starting with 20-50 simple tasks drawn from real failures and from work you currently check by hand.
Sources
- Anthropic — Demystifying evals for AI agents — Primary source for the terminology, graders, repeated trials, and starting with 20-50 tasks
- Anthropic — Building effective agents — Background on agent and workflow design