Verification discipline for coding agents

As of: 2026-09

The optimism problem

Agents are structurally optimistic: the model "sees" the fix working because the code reads correctly. Reading is not running. The gap between plausible and verified is where agent-made bugs live.

The verification ladder (each rung must pass before claiming done)

1. Reproduce first. Before fixing a bug, reproduce it. A fix for an unreproduced bug is a guess.
2. Fix.
3. Re-run the ORIGINAL repro. Not a new test you wrote that happens to pass — the original failing case.
4. Run the project gates. tsc, lint, build, test suite. Zero new failures.
5. Check side effects. Who else calls this code? What else changed behavior?

Honest reporting

Report what was verified and what was not: "Fixed X, verified via repro + full test suite. Could NOT verify the deployment path — needs a staging deploy." A scoped claim beats a confident lie.

Anti-patterns

• Tests that assert nothing (expect(true).toBe(true)) — written to make the suite green.
• "Verified" = "the code looks correct now".
• Deleting failing tests instead of fixing the cause.
• Marking a task complete because the token budget is running low.