Presentation: The Agent Harness: Control Planes, Invariants, and Approval Boundaries for Production AI Agents

OpenAI's Vinoth Govindarajan discusses the challenges of running AI agents in production environments, beyond just model hallucination. He presents the key principles of reliable agent harnesses, which include establishing explicit state ownership, serializing concurrent state mutations, scoping execution authority, and validating actions at the user-visible edge.

Govindarajan emphasizes the importance of these principles through real-world case studies, such as OpenClaw, where the user-visible delivery path appeared healthy, but the durable record behind it was incomplete, leading to a failure in reliably reconstructing the future for the next turn. He argues that silent success is worse than a crash, as it creates a lie where the system appears to have succeeded, but the operator has no reason to doubt that it lost part of its memory.

Govindarajan uses the example of an agent remembering a refund for a customer to illustrate this point, as the assistant said it would remember for the next turn, but the system could not reliably reconstruct the future for the next turn. He then introduces the concept of an agent harness, comparing it to a car's steering, brakes, dashboard, black box, and transmission.

The model, like the engine, is crucial, but the harness provides control and ensures the system operates reliably in production. Govindarajan stresses that ownership of state is essential, as a fact needs one owner and one replay path, and if the system cannot reliably reconstruct the fact later, it does not truly own the fact. He also highlights the importance of ordering mutations, preventing accidental interleaving, and proving actions, as the transcript and model intentions do not equate to the receipt of what was attempted, approved, and committed at the user-visible edge.