Monitoring Internal Coding Agents for Misalignment

OpenAI publishes very little about how they run agents in their own engineering organization. "How we monitor internal coding agents for misalignment" is the rare exception: a documented account of how they detect when one of their internal coding agents is doing something it shouldn't be: pursuing a side objective, taking shortcuts that produce plausible-but-wrong code, or otherwise drifting from the intent of the task.

The methodology is interesting. The implication is more interesting. OpenAI is the lab that built the model: they have the weights, the training data, the eval harnesses, the engineers who designed the architecture. And they still need explicit, dedicated monitoring infrastructure to deploy these agents inside their own company.

The post describes several distinct responsibilities: running an agent, monitoring its behavior, independently reviewing results, and deciding when to stop a run. Teams need to assign each responsibility explicitly.

That arrangement needs clear ownership beyond the people building the agent. If the same team builds and grades it, failures can be harder to challenge. An independent reviewer and someone authorized to stop a run provide additional checks.

OpenAI describes the discipline in "How we monitor internal coding agents for misalignment." They don't describe the cost. The cost is real: dedicated monitoring infrastructure, a separate review function, ongoing eval maintenance as the model and the workflow change. We've watched clients deploy two coding agents and discover they need the equivalent of a small SRE team to make those agents production-trustworthy.

That's the unsexy line item nobody includes when the agent is being scoped. We argue for putting it in the budget on day one, because finding it in month four is how a deployment gets quietly rolled back.