How autonomy works
Walk the canonical 11-step sequence a GreenLightz-governed AI agent runs when it proposes a consequential action, the environment misbehaves, and the outcome cannot be observed. The engine makes every branch automatically. No GreenLightz operator is on the path.
Step 1 of 11
The customer-facing AI agent proposes a $120 refund on behalf of a customer.
The AI agent posts a canonical AgentEvent (action_type=refund, amount_usd=120, tenant_id, actor_id, target_id, correlation_id). No dollar has moved. Nothing has executed. The proposal enters the deterministic disposition machine.
Each of the 11 transitions is a deterministic automatic disposition sourced from the tenant's versioned policy pack. The engine never invents an outcome, never widens the safe envelope, and never asks a GreenLightz operator to resolve a state that is defined in the supported space.
§XVII visual sequence · §XVIII allowed-before-proof