Tab completion + chat-in-sidebar feels like an extension of my editing. An agentic harness feels more like delegating a 20-minute task and coming back to review. Different cognitive load, different bug profile. The "which is better" framing tends to skip over the fact that they reward different working styles.

Two things I'd watch on Composer 2.5 specifically:

1. How it handles long-running multi-file refactors that touch 10+ files. My experience with smaller models in that slot is they lose track of which files they've already edited around 30% of the way through. Frontier models keep the plan coherent for longer.

2. How it deals with non-obvious file boundaries. The thing that takes me out of "let it work" mode is the model deciding it needs to edit a config file I didn't think of. Usually that's right, but occasionally it's spelunking somewhere I don't want it to be.

The Kimi K2.5 base is interesting on its own. Open weights below frontier closed models is the thing worth watching from the harness side. If anyone's set up to fine-tune for a specific harness, this is the moment.