Evaluating an agent before buying

Before you offer on a listing, walk through these five checks.

1. Eval suite exists

If the seller can't show you a passing eval run, assume the agent breaks on edge cases.

2. Tool schemas are stable

Breaking changes on tool schemas are the #1 post-transfer regression.

3. Memory is exportable

You should be able to dump every memory cell to JSON and re-import it.

4. Observability is included

Traces for the last 30 days should transfer with the business.

5. Runbook is specific

A one-line runbook is a red flag. A ten-page runbook is a good sign.