Stop Prompting Harder

When an agent stalls, inspect its tools, context, and feedback loops before repeating the prompt.

A coding agent surrounded by clear tools, tests, and feedback signals.

When a coding agent stalls, a louder prompt may not be the missing ingredient. OpenAI's Codex team reported that early progress was slow because the environment lacked useful tools, structure, and feedback paths. Their response was to build capabilities agents could find and use—not simply retry harder.

The method

Diagnose the missing capability; make it legible and enforceable; then observe the next run.

The source describes repository structure, tooling, observability, documentation, and iterative reviews. Its productivity figures are team-reported, not a controlled test of any one change. Our local comparison is a HYPOTHESIS , not an experiment already run.

Try one bounded feature: record a baseline trace, change one harness capability implicated by a failure, and compare correctness and human repair time. Keep the model, task, and rubric fixed. Stop if the baseline is not comparable.

Transfer rule: improve the harness when evidence points to a missing tool or feedback loop; avoid attributing gains to a change without a controlled comparison.