
Engineering article reviewResearch Insights Made Simple
Harness engineering: agent environments
Five months in which Codex wrote all of a product’s code while people built its environment
/ Research Insights Made Simple · Agent harnesses 3/7 · Harness engineering
Slide contents
1. Harness engineering: agent environments
Five months in which Codex wrote all of a product’s code while people built its environment
2. Five months without hand-written code
3. Humans steer; agents write everything
4. Agent failures signal a missing capability
5. Human QA became the bottleneck
6. Every change boots its own app
7. Codex queries logs, metrics and traces
8. What Codex can’t see doesn’t exist
9. AGENTS.md is a table of contents
10. Repo knowledge is versioned and machine-checked
11. Domain layers are enforced mechanically
12. Lint errors tell the agent the fix
13. Enforce invariants, not implementations
14. One prompt: from bug to merge
15. At high throughput, fixing beats waiting
16. Pay down debt daily, not on Fridays
17. The numbers show scale, not cause
18. Five months observed; years remain open
19. Measure each environment capability you add
20. Count accepted tasks and human time
21. Discipline moves from code to scaffolding