The 12-point audit for teams running AI coding agents in production.
After 6 months of ThumbGate telemetry across Claude Code, Cursor, and Codex installations, we identified the 12 failure patterns that cause 94% of all agent incidents. This checklist helps you find them in your setup before they cost you money.
- 1. Authorized paths: Agent can only edit files within the declared task scope
- 2. Protected files: Config files, .env, CI/CD scripts, and lock files require explicit approval
- 3. Branch guard: Agent cannot push to main/master without human review
- 4. Signature checking: Agent verifies function signatures against actual source before calling
- 5. Import validation: Agent checks that imports resolve before writing them
- 6. Test gate: Agent must run tests after each edit; failures block the commit
- 7. Config diffing: Any change to configuration triggers a diff review
- 8. Dependency lock: Package.json/requirements.txt changes require approval
- 9. Deployment wall: Agent cannot trigger deploys, migrations, or infra changes
- 10. Comment preservation: Agent must not delete existing comments
- 11. Format respect: Agent must not reformat code outside the edit region
- 12. Context persistence: Agent's task context survives session restarts
| Score | Risk Level | Action |
|---|---|---|
| 12/12 | Production-ready | Monitor and iterate |
| 9-11/12 | Moderate risk | Fix gaps within 1 week |
| 6-8/12 | High risk | Stop production deployments until fixed |
| <6/12 | Critical | Agents will cause incidents. Fix immediately |
I'll run this checklist against your actual codebase, identify every gap, and ship custom prevention rules for your stack.
$499 AI Agent Reliability Diagnostic →
Includes:
- Full codebase audit (any language, any stack)
- Custom ThumbGate prevention rules for your failure patterns
- CI/CD integration
- 30-day follow-up with metrics report
npx thumbgateStar the repo: https://github.com/IgorGanapolsky/ThumbGate