Slide contents

1. 1. Harness design for long-running (autonomous) application development

   Planner, generator, evaluator: what pays off, and what goes when the model improves

2. 2. An evaluator joined taste and correctness

3. 3. Every harness component encodes a model assumption

4. 4. On long tasks, context and self-assessment fail

5. 5. Resets trade continuity for a clean slate

6. 6. Tuning a skeptic beats tuning self-critique

7. 7. Four criteria turn taste into grades

8. 8. The last iteration is not always best

9. 9. Three agents split scope, build and acceptance

10. 10. “Done” is agreed before any code

11. 11. One criterion below threshold fails the sprint

12. 12. Full harness: over 20× the solo cost

13. 13. The solo build’s game did not work

14. 14. Evaluator findings name a likely cause

15. 15. Untuned, Claude is a poor QA agent

16. 16. Simplify one component at a time

17. 17. Each new model retired some scaffolding

18. 18. The evaluator pays off beyond solo reliability

19. 19. Building took 91% of the DAW budget

20. 20. QA caught display-only features

21. 21. What the evaluator cannot observe stays unchecked

22. 22. The harness space moves rather than shrinks

23. 23. Autonomy needed a separate action check

24. 24. State outlived the replaceable harness