Its own copy to work in

Think of it when two agents, or you and an agent, will change the same project at the same time.

In everyday life

It is like giving each student in an exam their own copy of the paper, so no answer lands on someone else’s sheet. They still share the room, the clock and the one pencil sharpener.

One case

Two agents work in one folder of a bank’s phone app at the same time. Agent A fixes the login screen while agent B renames a shared colors file. B’s rename lands in the middle of A’s test run, and A’s tests fail. A decides the failure is its own bug and undoes 40 lines of B’s work.

Its own copy to work in
The ideaLimit it
Its own copy to work in

Giving each agent its own copy of the project to change, such as a separate git worktree or a cloud machine. A worktree keeps its edits from colliding with yours or another agent’s; a separate machine or container can also keep ports, databases and keys apart.

Limit it, Steering

Every case

Agents in one folder share the files, the build output and the running servers. Each one sees changes it did not make and cannot tell whose they are, so failures appear that nobody can repeat. A separate worktreeA second working folder for the same git repo, on its own branch. It keeps the working files apart; it does not separate ports, databases or keys. per agent removes the shared files: each works on its own branch, the work meets only at a merge you control, and a bad attempt is deleted instead of untangled. A worktree does not separate ports, databases or keys, though: two agents can still clash over one test database, so give each its own, or use separate machines when that matters.

The word

Engineers call this

isolation, worktreeGiving each agent its own copy of the project to change, such as a separate git worktreeA second working folder for the same git repo, on its own branch. It keeps the working files apart; it does not separate ports, databases or keys. or a cloud machine. A worktree keeps its edits from colliding with yours or another agent’s; a separate machine or container can also keep ports, databases and keys apart.

Give each agent its own copy to work in, and its own database and ports when it runs the app. Combine the work later, on purpose, through a merge.

Where this idea leads

Skills that encode it

  • principle-separate-before-serializing-shared-state pstackEliminate the shared write target first: give each actor its own file, branch or state directory.
  • arena pstackGives each parallel candidate its own output location, a git worktree where possible.
  • swarm pstackGives each parallel worker its own writable output when it writes.