Very interesting paper on making coding agents actually finish the job
(Bookmark this and read the paper when you have time)
They trained the model on 1,700 tasks where it gets points for every requirement it meets and zero if it breaks something that already worked
Like a kid who already knows how to build with Lego, but learns to check the full instructions before saying it's done
Model: Kimi K2.7 Code
Harnesses:
- Claude Code
- mini-swe-agent
- Terminus 2
Benchmarks:
- SWE-Bench Pro
- DeepSWE
- Terminal-Bench
- SWE-Marathon
It improved on all of them, It didn't learn more code, it learned to finish