Plan mode

Plan mode is an agent mode in which the agent can read files, search the codebase and run harmless commands, but cannot edit or execute anything that changes state. Its job is to come back with a plan: which files it will touch, in what order, what it is unsure about, and how it will verify the result. You read the plan, correct it, and only then switch to an editing mode. The permission mode underneath is 'ask for everything that mutates', with a system-prompt nudge to investigate first.

A wrong plan is cheap. A wrong implementation is not: by the time you notice, the agent has changed fifteen files, the tests are green for the wrong reasons, and unpicking it costs more than starting over. Plan mode moves the correction to the point where it is a sentence instead of a revert. It also surfaces the assumptions the agent would otherwise silently make about your codebase, which is where most AFK runs go wrong.

Use it for anything bigger than a one-file change, and pair it with a written spec so the plan has something to be checked against. A plan that survives a round of grilling becomes a good handoff artifact: save it to a file, start a fresh session, and let that session execute with the sharp end of its context window free for the work. Skip plan mode for trivial edits; the overhead isn't worth it there.