Concurrent subagents editing the same file silently clobber each other — enforce file ownership, and re-partition the moment scopes overlap

Multi-agent coding has no file locking: each subagent edits its own view and writes the whole file, so two agents editing the same file = last-writer-wins and the other agent’s changes vanish silently (no merge, no conflict marker). Mitigations that worked: (1) Partition by FILE/DIR ownership up front — one agent owns each file; parallelize only across disjoint files. (2) The hazard reappears mid-flight when you PROMOTE a read-only reviewer to an implementer: a design-review agent that was only screenshotting got user-directed to start editing the exact UI files a bug-fix agent was already editing. The fix is immediate re-partitioning: pick ONE owner per contested file, message the other agent to STOP touching those files AND to report exactly what it already changed there so you can hand that diff to the new owner (don’t assume nothing was written yet). (3) Split bugs by file, not by theme: a ‘listener leak in home.js’ and a ‘nav cleanup in nav.js’ should go to whoever owns those files (the design agent), even though they’re ‘logic’ fixes, so the bug-fixer never has to touch UI files. (4) As coordinator, after concurrent passes land, reconcile: git diff + run tests + re-screenshot to confirm neither agent’s work was lost before committing a coherent tree.