Slide contents

1. 1. SWE-agent: interfaces change outcomes

   How search, editing and feedback help the same model

2. 2. A model needs its own IDE-like interface

3. 3. The ACI shapes both command and feedback

4. 4. Resolved means the repository tests pass

5. 5. SWE-agent resolves 12.47% of full SWE-bench

6. 6. The ACI adds 7 pp over Shell-only

7. 7. Good search forces a narrower query

8. 8. Iterative search is worse than none

9. 9. The file window keeps the agent oriented

10. 10. One edit returns the updated state

11. 11. A rejected edit must enable recovery

12. 12. The editor contributes more than linting

13. 13. Old output becomes a placeholder; actions stay

14. 14. Maximum context loses to selection here

15. 15. Successful runs reproduce and locate first

16. 16. Agents succeed quickly and fail slowly

17. 17. Six attempts expand the solvable task set

18. 18. Half of failures: wrong or narrow fixes

19. 19. Transfer the method, not the constants

20. 20. Design actions, measure whole tasks