Risk-based autonomy means the amount of work an agent may finish without a person depends on how risky and reversible the change is, and grows only as verification and rollback prove trustworthy. All three sources that discuss it keep production behind a human gate.
Sizing by path (GH-600)
| Risk | Example paths | Design |
|---|---|---|
| Low | docs/, formatting | Auto-merge after required checks |
| Medium | src/, dependency bumps | PR, checks, at least one review |
| High | infra/, .github/workflows/ | CODEOWNERS, multiple reviews, stricter rulesets |
| Critical | Production deploys, settings, secrets | Environment approval: the agent prepares but does not execute |
As described in module 2.1 A GitHub Actions environment with required reviewers pauses any job targeting it until a person approves.1
Widening in steps (AI-native SDLC playbook)
- Start with read-only work such as build-failure triage and changelog drafts.
- Let agents propose changes only through pull requests and branch protection.
- Run jobs in sandboxes with short-lived, scoped credentials.
- Expose deployment and rollback as allowlisted tools per environment.
- Allow more autonomy in development than in production.
- Keep the production gate human and rehearse rollback.
Rolling out Claude Code Auto Mode
The Auto Mode guidance follows the same shape: start narrow, keep explicit deny and ask rules, watch what is denied, widen gradually, and keep human review for production infrastructure.3
Reliability assumptions
GH-600 adds that agent workflows should assume failure: bounded retries for transient check failures, escalation to a person after a check fails twice (with what failed, what was tried, and a suggested next step), and rollback readiness for high-risk changes.1
Related
- Agents propose, people and policy accept
- Least-privilege tool access
- Source: GitHub Certified: Agentic AI Developer
- Source: The AI-Native SDLC Playbook
- Source: How Claude Code Auto Mode Works