AI Automation Governance Framework
Agent classification, approval gates, monitoring, incident response.
Governs design, deployment, and operation of all AI agents. Four classes, four approval gates. Every agent needs: brief, data flow, tool manifest, permission model, fallback rules, rollback plan, test plan. Kill switch mandatory on every deployed agent.
Key concepts
- Class 1 — Read Only: reads data, generates output, no external action
- Class 2 — Notify: reads data, sends notifications
- Class 3 — Write: reads and writes to systems (claude-code@mandantis.com is Class 3)
- Class 4 — Orchestrate: spawns sub-agents, chains multi-system workflows
- Kill switch: Cloudflare KV · namespace: mandantis-platform · key: agent:[agent-id]:enabled
- Claude Code is AGT-001 — Class 3. Kill switch tested 2026-04-20.
- Max turns default: 20. Override requires CLAUDE.md justification.
- Every agent must have: register entry, tool manifest, kill switch, test plan, monitoring