Guardrails bound what an agent may do without asking. Each action passes a policy check, and risky actions pause for a person.
Tier | Agent may | Example in chip work |
|---|---|---|
Read | Read named sources | Read |
Propose | Draft text with no side effects | Draft a fix plan for review. |
Write to scratch | Write inside one scratch directory | Write |
Write to repo | Change tracked sources after approval | Edit |
Run jobs | Start named jobs after approval | Start the |
Lesson | What it covers |
|---|---|
The five tiers and a runnable policy checker. | |
Containment for runs and gates for risky actions. | |
Budgets with worked arithmetic and per-action logs. | |
What counts as a secret and how to keep it out. | |
Injected instructions in content and limits on damage. |
Write to repo. It may not proceed alone; that tier needs a recorded human approval first.