Slide contents

1. 1. Long-running agents: work handoffs

   What the next session receives: requirements, verified results and the next step

2. 2. The next shift needs verified state

3. 3. Every session starts without memory

4. 4. Compacted conversation is not working state

5. 5. An oversized step breaks off halfway

6. 6. Visible progress does not prove completion

7. 7. The first session sets up the rest

8. 8. Four artifacts answer four questions

9. 9. Features are written as observable scenarios

10. 10. Status changes only after behaviour is verified

11. 11. The JSON observation needs testing

12. 12. Check the foundation before building further

13. 13. One feature gives a verifiable step boundary

14. 14. API success leaves the UI unverified

15. 15. Verification depends on what tools expose

16. 16. Commits preserve recovery points

17. 17. A useful record ties results to evidence

18. 18. Both roles answer every failure

19. 19. Engineering experience supplies measurable hypotheses

20. 20. We propose testing harnesses with forced resets

21. 21. Long-running work is built from verifiable steps

22. 22. Long-running work became a team effort

23. 23. Context resets dropped, separate evaluator added