OpenAI Engineer Says Production Agents Fail on State Control, Not Hallucination

Production AI agents break for reasons that have little to do with the model making things up, according to Vinoth Govindarajan of OpenAI. In a presentation published by InfoQ, he argues that the failures that matter in deployed systems come from how agents own state, mutate it, and exercise authority — not from hallucination alone. The talk lays out a set of engineering principles for what he calls the agent harness, the layer that surrounds a model and turns it into something a business can run.

Govindarajan's first principle is explicit state ownership: every piece of state an agent touches should have a named owner, so it is clear which component reads it, which writes it, and which is allowed to change it. The second is serializing concurrent state mutations, a familiar problem from distributed systems that becomes acute when multiple agent steps or tool calls try to update the same record at once. Without that ordering, agents can overwrite each other's work in ways that look like reasoning errors but are actually race conditions.

The third principle concerns scoping execution authority. Rather than granting an agent blanket access to tools or APIs, the harness limits what each step can do to the minimum needed for that step, which caps the blast radius when something goes wrong. The fourth is validating actions at the user-visible edge — checking the result where a person will actually see or act on it, instead of trusting an internal check deep in the pipeline. Together these form what Govindarajan describes as a control plane for agents, distinct from the model itself.

The talk leans on real deployments, including the OpenClaw case, to show how these failures surface in practice. The pattern he highlights is that teams often ship an agent, watch it behave well in demos, then discover that production traffic exposes state conflicts and over-broad permissions that no amount of prompt tuning fixes. That framing puts the harness, not the model, at the center of reliability work for agent products.

For teams building agents, the message is that the hard engineering is in the surrounding system: who owns state, how writes are ordered, what authority each action carries, and where validation happens. Those choices are largely independent of which model is under the hood, which means they can be designed and tested before a single prompt is written. It also suggests that agent reliability will be judged by the same standards as any other distributed system.

What remains open is how much of this harness work can be standardized. If control planes, invariants, and approval boundaries can be shared across vendors, agent development gets cheaper and safer; if each team rebuilds them, reliability stays a bespoke effort. Govindarajan's principles are a starting point, but the field has yet to agree on what a default agent harness should look like.