Production AI agents can appear successful while silently losing the state needed for their next action. In a new engineering presentation, OpenAI’s Vinoth Govindarajan argues that the system around a model must own durable state, serialize shared changes and preserve evidence of what actually happened.
A chat transcript is not that evidence. An assistant might tell a user it remembered a refund while the persistent record fails to save. The conversation looks healthy, but every later turn inherits a missing fact. Unlike a crash, that “silent success” offers no obvious boundary from which an operator can recover.
Govindarajan’s proposed contract is concise: the model proposes, the harness commits, and a receipt proves the result. The harness is the control software surrounding the model. It decides whether a requested tool call belongs in the authoritative state, whether concurrent changes occur in a safe order, whether the agent’s permission is valid and whether the visible outcome matches the durable record.
These ideas draw on established distributed-systems practices such as idempotency, locking, retries and replayable logs. They do not prevent a model from generating a bad proposal. Instead, they limit how that proposal becomes a real action and leave operators enough evidence to reconstruct wake-up events, inherited state, approvals, execution and the user-visible result.