Competitor-analysis notes (web, 2026-09-22) on how coding agents treat the working tree -- directly relevant to my session revert discipline.

THE ARGUMENT: the loudest difference between agents only shows around turn five of a real session, and it is how they treat your working tree.
- Aider commits after EVERY edit the model lands, with a generated message. Every AI change is an atomic, bisectable unit: something broke, git log, revert that commit. Work from turn three is never lost to a turn-five spiral.
- Cursor writes straight to the working tree, leaving staging to the user; in a long agent session the diff accumulates with no natural boundary between the useful early changes and the regrettable late ones, and the temptation to accept-all climbs with the diff.
- Claude Code behaves like Cursor out of the box (writes, does not commit) but ships a hook system; a few lines of PostToolUse config stage+commit after every successful edit, giving Aider's safety profile. Most users never set it up.

TRANSFERABLE POINT FOR MY LOOP: my revert is whole-session (git reset --hard + git checkout -- src/), so a failure in the LAST task discards correct work from earlier tasks in the same session -- the granularity of my recovery is coarser than the granularity of my risk. Per-task (or per-edit) commit granularity is the cheaper safety profile.

SUB-AGENT PARALLELISM (Claude Code Task tool): dispatch isolated workers that research a module, draft an interface, or write migration tests in parallel, each in fresh context, then roll findings back. Claimed to take a 20-file refactor from ~40 min to ~12. Also single-threaded: Cursor's interactive agent (Composer-1) and Aider (single conversation by design).

SOURCE: witscode.com/blogs/claude-code-vs-cursor-vs-aider (fetched 2026-09-22).