What is an agent workspace?

The definition

An agent workspace is a place to run several coding agents at once and review what they produce. It is not an editor and it is not a terminal emulator: it is the layer that manages where each agent works, what it is allowed to touch, and how its output gets back into your branch.

The unit it is organised around is a task: an instruction, a checkout, a branch, a diff, and a decision to merge or discard.

The four parts

• Task. An instruction with a definition of done — a test that has to pass, a command that has to exit zero.
• Checkout. A git worktree per task, so two agents cannot edit the same file and neither can disturb your working copy.
• Terminal. A real terminal per session, because the agents are CLIs and the escape hatch has to stay open.
• Review. A diff per worktree and a merge queue, so work is read before it lands rather than discovered afterwards.

How it differs from the neighbours

Compared with an AI IDE, the difference is the unit: an IDE optimises the loop where you and a model edit one file together. A workspace optimises the loop where you hand off, walk away, and come back to a proposal.

Compared with a smart terminal, the difference is the structure around the shell. The terminal is how an agent gets things done; it is not the thing you are managing. What you are managing is tasks and diffs.

And compared with doing it all by hand — worktrees, branches, diffs, merges — the difference is that the bookkeeping is done for you at three sessions, which is exactly where doing it by hand stops being fun.