The Swarm: Parallel Coding Agents in Git Worktrees

Without isolation, parallel agents corrupt each other's work in ways that pass silently and reach production.

Recognizing the real bottleneck tells you to change scheduling, not chase a faster model.

The failure arrives as corrupted work, not an error, so you must prevent it structurally.

Keeping the orchestrator out of the code preserves the one global-view role that cannot be isolated.

You get real parallelism cheaply, without re-cloning history for each agent.

Wrong authorship destroys your only record of which agent did what, and it fails without an error.

Cheap isolation is what makes fanning work out across many agents practical.

Budget your effort for merging, because that is where the swarm's cost concentrates.

The cheap-swarm trick can quietly destroy files that were never in the worktree.

A short rehearsal prevents silent bugs from reaching every learner's machine.

Independence at this step is what determines whether the whole night is cheap or expensive.

Per-process identity fixes the silent authorship race and lets you audit each agent's work.

Editing collapses your global oversight into just another isolated worker's blind spot.

Serial fan-in preserves attribution and pins every conflict to a single responsible branch.

Semantic conflicts pass every merge check and only tests after the last merge reveal them.

The command order is the only thing between routine cleanup and irreversible data loss.

Measured trials show fanning out can slow you down 39 percent when tasks overlap.

These four artifacts turn a single-agent queue into a repeatable parallel workflow.

Getting these right converts a slow single-agent queue into safe, auditable parallel progress.