A while back I wrote about parallelism—taking one big task, splitting it, and fanning it out to several agents at once. But I left a thread hanging: when several agents edit code at the same time, what keeps them from clobbering each other? That's what this post is about.
Let me start with the wall I ran into myself. The subagents an AI spins up on its own within a session—those are actually fine; it works out isolation by itself, I don't have to worry much. What burned me was the other kind: I manually opened several AI terminals and had them edit the same project at once. Several hands reaching into the same working directory, you write a bit, I write a bit, files overwriting each other, state in a mess—halfway through, nobody's work is clean.
What I want, and one boundary
What I want isn't complicated: several parallel hands, each doing its own thing, editing the same project without clobbering each other.
But let me draw a boundary up front—not every kind of parallelism needs isolation. If I send several agents out—one to dig through logs, one to analyze a cause, one to flip through docs—those are all read-only, and they can crowd into the same workspace with zero trouble; no need to carve out a plot for each. What actually needs isolation is the parallelism that writes to files. Reads can share, writes must split—that's the starting point for everything below.
