OpenAI agent beats top coders in 34-hour loop test

The contest result is the entry point rather than the conclusion. The improvement loop described — implement, score, then branch from the best-scoring candidates — kept producing gains until late in a 34-hour limit, which is what made the agent's win over top human competitors notable. The focus then pivots from that demonstration to the workplace, where the same loop's usefulness is tied to how deliverables get evaluated.

The piece is the second in a series, and its stated job is to lay out the three loops that run inside AI agents before narrowing to the improvement loop, which it says is spreading in use. The open question it poses is not whether the loop works in a contest but how far it carries into business, and where the walls sit. Evaluation of output is presented as the hinge for that transition, making scoring criteria for deliverables the practical bottleneck rather than the agent's ability to generate and revise candidates.