DNA / Bounded Autonomy
Bounded Autonomy
AI executes. Human decides.
Bounded autonomy is the principle that AI agents should operate within clearly defined boundaries, with human decision points at every meaningful threshold. Not because AI can't be trusted — but because accountability requires a human in the loop.
Key Principles
- 01HITL gates: Human-in-the-Loop at every production boundary, not just critical failures
- 02Zasada 11: Zero autonomous production deploys. Every deploy requires explicit human approval
- 03Kill switches: Every agent has a documented way to halt, roll back, or override
- 04Audit trails: Every agent action is logged with handoff docs — reconstructable at any point
Real Examples from Production
Agent OS Reviewer (HITL)
The LangGraph pipeline pauses at the Reviewer node. Code cannot proceed to Summarizer without explicit human approval via the Mission Control UI.
agent-os/src/graph.py →
Zasada 11 enforcement
No agent in the ecosystem has deploy credentials. GitHub Actions workflows require manual trigger. This is not a limitation — it is architecture.
WorkFlow/AGENTS.md →
GHA manual gates (ZZPackage)
WordPress/WooCommerce deploys require manual workflow dispatch. GUARD_01-04 validates before any deploy is even permitted.
zzpackage.flexgrafik.nl/AGENTS.md →
Traditional vs AI Systems Architect
Traditional Developer
AI agent runs autonomously, deploys on its own, discovers the bug 3 hours later in production.AI Systems Architect
AI agent prepares the deploy. Human reviews the diff. Human triggers the deploy. AI cannot override this.