GUIDE-001 · CHECKED
How to verify AI-generated code before merge
Verifying AI-generated code means checking what the agent actually ran, not just what it reported. Read the run record, compare claims to recorded output, confirm which checks ran, and keep the merge decision with a person. A passing run is not proof of correctness.
§ 01
Steps before merge
- Get the run record. Read what the agent received, the model, the ordered tool steps and the final status. If all you have is the agent's chat, you have claims, not a record.
- Compare claims to recorded output. "Tests pass" is a claim; look for the recorded test command and its exit status in the run record.
- Check what was verified. Which checks actually ran: tests, lint, typecheck, build. A missing check stays unverified regardless of the agent's summary.
- Read the diff yourself. The run record shows what the agent did; the diff shows what it changed. Both are needed.
- Decide with a person. Verification ends in a human merge decision, not in a green checkmark from the agent.
§ 02
Signals to look for
- Recorded commands with real exit statuses, not narrative summaries.
- Explicit marks for checks that were skipped or never run.
- Artifacts you can open: logs, exports, changed-file lists.
- A status that distinguishes "the runner failed" from "the candidate failed".
§ 03
Why the run record is the evidence
A chat transcript is what the agent reported. A run record is what the infrastructure observed: the actual tool calls, their outputs and the recorded status. When the claim and the record disagree, the record is the evidence to inspect.
AdaptOrch keeps hosted run records for agent work over MCP and the API. See what an agent run record contains and the patch verification guide for the diff-side checks.
§ 04
What verification does not prove
A passing check covers only what it executed. Verification tells you what ran and what its outcome was; it does not prove the change is correct, complete or safe. The claim boundary is explicit about this.
§ 05
FAQ
Q.01How do I verify AI-generated code?
Read the agent run record for what actually ran, compare the agent's claims to recorded outputs, confirm which checks ran, review the diff, and keep the merge decision with a person.
Q.02Is a passing test enough?
Only for what that test covered. A pass shows a specific check ran and returned green; it does not prove the change is correct or complete.
Q.03What is the difference between a transcript and a run record?
A transcript is what the agent said. A run record is what the infrastructure recorded: tool calls, outputs and status, kept separately from the agent's own report.
NEXT
Try it on one agent change.
Connect your coding agent, run one change through it, and read the record next to your usual review. Starter is free.
Starter is free with 50 admitted runs a month. Pro is $39 a month for 500 admitted runs and 5 workspace users. Team is $149 a month for 2,500 admitted runs and 15 workspace users. Compare plans · Start free on Starter
Connect from Claude Code, Codex CLI or Cursor · Read the AI PR review guide