Slide contents

1. 1. SWE-agent: interfaces change outcomes

   How search, editing and feedback help the same model

2. 2. The hard exam came first

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

4. 4. The ACI shapes both command and feedback

5. 5. Resolved means the repository tests pass

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

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

8. 8. Good search forces a narrower query

9. 9. Iterative search is worse than none

10. 10. The file window keeps the agent oriented

11. 11. One edit returns the updated state

12. 12. A rejected edit must enable recovery

13. 13. The editor contributes more than linting

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

15. 15. Maximum context loses to selection here

16. 16. Successful runs reproduce and locate first

17. 17. Agents succeed quickly and fail slowly

18. 18. Six attempts expand the solvable task set

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

20. 20. Transfer the method, not the constants

21. 21. Design actions, measure whole tasks

22. 22. SWE-smith scales up training tasks

23. 23. mini-SWE-agent keeps just bash

24. 24. SWE-bench Verified also needed an audit