Agents and branches

Selecting cards on the Board and pressing Plan doesn’t start work by itself — it asks your own Claude to propose how to split the selection into agents. Nothing runs until you decide.

What an agent is

An agent is one CLI (Claude Code, Codex or Antigravity), one isolated worktree, one branch, and an ordered list of cards. It works through its cards one at a time, in order — one card’s prompt goes out, and the next is typed only once that turn lands.

Planning and accepting

Pressing Plan writes a proposal: a read-only pass over your selection groups the cards and suggests names and branches. You see the proposed groups before anything exists — edit them, split them, or leave them as suggested. Accepting is per group: each group you accept becomes a real agent, with its own worktree and branch, and starts on the machine’s next check-in. A group you haven’t accepted yet does nothing, and a group waiting on a sibling you haven’t started yet says so.

Branches

Each agent works on its own branch, named agent/<name>, cut from your base branch. You can see it, and how far ahead it is, on the agent’s own page and on the Board. Nothing touches your base branch until you approve and merge the agent’s work — or accept its mockup or write-up, for a design or research card.

Where an agent shows up

An agent’s cards render in whichever column matches the agent’s own state — Working while it’s running, Stuck if it needs you, Review once it’s delivered everything. See Columns and what they mean. A card an agent is holding isn’t selectable on its own; the agent is the thing you check on, act on, and eventually review.