Orchestrating coding agents in parallel

Running a single coding agent is straightforward. Running ten at once is where things get interesting — and where most setups fall apart.

The naive approach is to point every agent at the same working directory. That breaks immediately: agents overwrite each other's files, tests run against half-applied changes, and you lose any ability to reason about what happened.

emdash gives each agent its own Git worktree. The agents share history but never share a working tree, so their changes stay completely isolated until you choose to merge. The dashboard then shows every run side by side, with a live diff for each one.

The result is that parallelism becomes a coordination problem you can actually see and control, rather than a race condition you hope to avoid.