Trust Levels per Agent Action
Governance pattern documented by intodesignsystems.com: graduating trust granted to AI agents by action type, not all-or-nothing. Examples: auto-merge OK to add a unit test, draft-PR only to modify a DS component, suggest-only to touch tokens. GitHub Primer goes as far as: agents authorized to create issues only, never to merge. Fine granularity avoids the 'all-AI or nothing' trap. Detailed implementation (see pattern `per-action-trust-levels`): auto-merge if tests pass and action is safe (typo, missing story, lint fix), draft-PR if tests pass and action is structuring (token modified, variant added, props added) with human review expected, suggest-only for critical actions (atomic component, layout pattern, a11y), the agent can only propose a comment on an existing PR. The same reasoning holds outside code: a prospecting agent like Harvey finds businesses, reads their sites and checks addresses with no approval at all, because a mistake there costs nothing, and asks for your go-ahead only at the one irreversible spot, sending the email. Trust is graded by the cost of the mistake, not by how hard the task is.
Strengths
- Exits the auto-merge / no-AI extreme, nuanced, defensible, calibrated policy
- Compatible with glass-box-ai: high trust-levels → detailed surfaced decisions required
- GitHub Primer model proves enterprise viability: agents = issues only, not merge
Limitations
- Requires initial categorization effort: which AI actions in which risk categories?
- Limited 2026 tooling: most CI/CD have no native granularity, must code manually
Best for
- Design system manager presenting an AI policy to management or governance committee
- Team scaling AI to 20+ devs wanting to avoid both extremes (full-auto vs full-block)