What is an AI agent eval?
An eval is a test harness for an agent. A team writes tasks that resemble real work, runs the agent against them, and scores the results: task completed or not, output quality, tool choices, cost, and failure modes. Run the same eval after a model swap or a prompt change and the scores show whether the change helped or hurt.
- Before deployment: is this agent good enough to trust with real work at all?
- During changes: did the new version get better, or quietly worse, on the same tasks?
- During deployment: sampled checks on live traffic can flag drift and new failure patterns.
All of that is estimation. An eval tells you how an agent tends to perform across the tasks you sampled. It cannot tell you what happened on one specific job for one specific client. No score can.
Evals vs verification: two different questions
| Question | Eval | Receipt verification |
|---|---|---|
| What does it measure? | Quality across sampled test tasks | One real call and its saved result |
| When does it run? | Before deployment, and around changes | On every production call, as it happens |
| What does a good result mean? | The agent looks ready, on those samples | The saved result was not changed since it was saved |
| Who can check it? | Mostly the team that built the eval | Anyone with the receipt link, free, no account |
Scope, stated once: an agent receipt is a record of one ai agent tool call that anyone can check: it keeps the exact saved bytes and their sha-256 digest, and it proves the result was not changed since it was saved, not that it was right.
Read the table carefully and the two tools stop competing. An eval estimates quality on samples. A receipt proves what one real call saved and that it was not changed after. A team that only evals has a forecast and no record. A team that only collects receipts has a record and no forecast. Serious production work wants both.
Worked example: a receipt from live work, not a test task
From the public stream on October 25, 2026: a capability_search call saved at 2026-10-25T18:35:28Z, receipt at zambo.dev/run/48493b66-f9bf-4bff-80cb-052af5716698. This was not an eval task. It was a real call, run by a real session, doing real lookup work.
The verification endpoint returned verified for this receipt. An eval could tell you how capability_search tends to perform across a hundred test prompts. Only the receipt can tell you what this call returned, and let a second person confirm the record is intact at zambo.dev/verify. Zambo issues a receipt on every call in the AER-1 format, specified by the IETF Internet-Draft draft-zambo-aer1-26, across one MCP endpoint at https://zambo.dev/mcp, 153 tools total, 35 on the default surface.
How to run both halves
- Eval before you ship. Build tasks from the real jobs the agent will do, score them, and keep the eval as your regression check.
- Receipt every production call. Add Zambo at zambo.dev/add so live work leaves a checkable record by default, not by request.
- Feed production back into the eval. When a receipt shows a call that went wrong, turn that job into a new eval task. Production finds the cases your samples missed.
- Answer each question with the right tool. Readiness questions go to the eval. What happened questions go to the receipt.
Getting started costs nothing: 20 free calls per day, each with its own receipt. A Day Pass is $0.99 USDC on Base for 24 hours of full access. The Zambo Pass is $29 for 30 days, one payment, with an annual option at $290.
Questions, answered
What is an AI agent eval?
An AI agent eval is a structured test that scores an agent against a set of tasks, prompts, or scenarios, before or during deployment. It estimates how well the agent performs on samples, so a team can compare versions, catch regressions, and decide whether the agent is ready for real work.
What is the difference between an eval and verification?
An eval estimates quality on samples: it scores how an agent did across test tasks. Verification checks one real record: a receipt proves what one actual call saved and that the saved result was not changed after. One is a forecast from samples, the other is a check on a specific piece of production work.
Do evals replace receipts, or receipts replace evals?
Neither. Evals belong before and around deployment, where a team is choosing and tuning an agent. Receipts belong in production, where a specific client, user, or auditor asks what happened on a specific job. A strong setup uses both, each for the question it can actually answer.
What does a receipt prove, in one line?
A receipt proves the saved result was not changed since it was saved, not that it was right. It keeps the exact saved bytes and their sha-256 digest, so anyone can recompute the digest and check the record at zambo.dev/verify.
How do I start getting receipts on live agent work?
Add Zambo to your AI at zambo.dev/add. You get 20 free calls per day, each with its own receipt. A Day Pass is $0.99 USDC on Base for 24 hours of full access. The Zambo Pass is $29 for 30 days, one payment, with an annual option at $290.