In a multi-agent coding workflow, one agent plans and the others build. You give the lead a goal. It posts a task board, with each task naming the files it owns and what it waits on. Free agents take ready tasks, and a different agent checks each finished task before it counts. This is how it works in Friday.
What is the shape of the workflow?
- Add the agents you want to the session, such as Claude Code, Codex and Cursor.
- Make one the lead from its tab’s menu, with “Make it lead the others”. An owner or editor sets it, and only one agent leads at a time.
- Give the lead one goal, in the All agents view or its own lane.
- The lead splits the work and posts the board.
- Free agents take ready tasks. Another agent reviews each finished task.
- The lead hears once, when the board is through, and reports what was done.
What goes on the task board?
Each task has a title, what to do, the files it owns, the tasks it waits on and the agent it suits. The lead is told to cut work into pieces that do not share files and to hand out every piece up front, so they run at once, and to do a piece of its own meanwhile.
| Field | What it does |
|---|---|
| Files it owns | Keeps two agents from editing the same file |
| Waits on | Holds a task until the ones it depends on are done |
| Suits | Names the agent best placed for it, which gets first call |
| Status | Shown live on one card: who has each task, who is checking it |
A task is ready when everything it waits on is done and nobody has it. A busy agent picks up the next ready task by itself when it finishes, so nobody dispatches by hand.
How does the work get checked?
A finished task goes to a free agent that did not do it. That agent reads the change, runs the tests and ends with “Approved” or “Changes needed”. If approved, the task is done and what waited on it can start. If not, it goes back to its agent, and after a third failed review it is blocked for the lead or a person.
Each report ends with facts from the session record rather than the agent’s own account: files changed with additions and deletions, the tests and checks run and how they came out, any commit, and how long it took. A task that changed no files, or with no other agent to check it, passes with a note.
How do the agents stay in step?
- A task names the other open tasks and who has them, so agents settle what they share, such as an interface, by asking each other.
- Agents say which files they will change before they write, and Friday warns when lanes in separate copies change the same one.
- An agent that stops short of its task is asked to carry on, up to five times in a run, unless it is waiting on a person.
- When a coding tool runs out of usage, its task is handed on to another agent that can take it, or resumes when the limit resets.
- A failed or blocked task reaches the lead at once, so it can unblock it while the others carry on.
The messaging underneath is covered in Claude Code and Codex talking to each other.
Where do people come in?
The board is not unattended automation. Everyone in the session sees the board and every agent’s steps, and a teammate can suggest a redirect or take control of an agent. Plan, Manual and Auto still decide what each agent may do without asking, and risky commands wait for a yes. A person reads the diff and accepts it. Friday ranks each run’s changes by risk, so the small safe ones can be cleared together.
When is this too much?
For a change that fits in one agent’s head, a lead and a board is overhead. Use it when a goal splits into three or more pieces that touch different files. Friday allows one writer per folder on a computer, so give each agent that edits its own worktree. Lanes in separate copies still get a warning when they touch the same file. Your usage limits matter as well: several agents on one plan reach the limit sooner than one.
If you are still choosing a tool for this, see how to choose an agent orchestrator.