A lost response can look like failure even when the action happened.
AI EXECUTION
AI can decide what to do.
It should not decide what happened.
When agents mutate real systems, model confidence is not execution truth.
AGENT
action intent →TRUTH ENGINE
→ governed executionPROVIDER
evidence ↩
SATISFIED · NOT_SATISFIED · UNKNOWN
WHY THIS MATTERS
Agents act faster than humans. Ambiguity still exists.
Acknowledgement, timeout and final occurrence are different facts.
Inference is useful for intent. It is not authority for external reality.
EXAMPLE · REFUND
The dangerous moment is after the request leaves the agent.
AgentRefund customer
↓ProviderRefund may execute
↓NetworkResponse lost
↓Truth EngineUNKNOWN
↓Truth EngineReconcile evidence
↓OutcomeSATISFIED / NOT_SATISFIED / UNKNOWN
The agent may request.Truth Engine governs execution authority and outcome truth.
ROLE SEPARATION
Agent intelligence and execution truth are different jobs.
WHERE IT APPLIES
Any agent action with a consequential external effect.
Issue refundCreate orderBook resourceProvision accountChange accessTrigger transaction
Truth Engine is not an agent framework or planner. AI agents are a client class. Current external evaluation is developer-controlled; production agent-framework adapters are not claimed yet.
EVALUATE THE BOUNDARY