GUIDE-003 · CHECKED

AI agent runner health is part of code review evidence.

When an AI coding agent reports a failed task, first ask whether the runner failed or the candidate failed. The distinction changes what a reviewer can conclude and what must be rerun.

Audience
Platform teams and reviewers of coding-agent runs
Checked
2026-11-19
Boundary
Not a correctness certificate

§ 01

Two failures that look alike

A runner failure means the environment could not complete the requested work: a tool was unavailable, a command timed out, authentication failed or the process crashed. A candidate failure means the requested check ran and found a problem in the generated code.

The agent run record should make the distinction inspectable through ordered steps, outputs and final status. If the record cannot show which command ran, treat the result as a verification gap.

§ 02

Signals of runner health

  • Tool and command steps have recorded outputs and exit statuses.
  • Timeouts, authentication errors and unavailable binaries are separated from test failures.
  • The final status explains whether the run completed, failed, degraded or was cancelled.
  • Skipped checks are visible instead of being folded into a success summary.

These signals describe the execution environment. They do not establish that the changed code is correct or safe.

§ 03

Triage the next action

  1. Runner failure: repair or rerun the environment step before judging the patch.
  2. Candidate failure: inspect the failing output and the changed code.
  3. Verification gap: run the missing check through an approved path or record it as unresolved.

Pair this triage with the AI-generated code verification guide. The goal is not to turn runner health into a badge; it is to make the remaining uncertainty visible.

§ 04

FAQ

Q.01What is an AI agent runner failure?

It is a failure of the execution environment or tool path, such as a timeout, missing binary, authentication error or process crash.

Q.02What is a candidate failure?

It is a problem reported by a check that actually ran against the generated change.

Q.03Does runner health prove the code is correct?

No. It only tells you whether the execution path produced inspectable evidence. Code review and relevant checks are still required.

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.

START FREE ON STARTER

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