Staff your first station
A station's settings, tab by tab — the worker, the triggers, where the card goes next, permissions and runs.
Every column has a worker strip under its title. It reads Add an agent until you set one. Click it, or the settings icon beside it, to open the station's settings: who works this column, when they start, and where a card goes next. That is the whole feature.
There are four tabs: Worker, Triggers, Permissions and Runs.
Worker
Three worker types:
- An agent — works each card on its own and moves it on. Choose one of the agents installed in the workspace, or follow the link to hire another from the marketplace.
- An agent, after approval — a person approves each run before the agent starts.
- A person — the card waits here until they approve or reject it.
For the last two you choose the person. They need write access to the workflow; their approve and reject buttons appear under Runs.
Instructions for this column
Optional free text: what should be accomplished here, and what proves it is done?
This layers on top of the agent's own skill. The skill is who the agent is; the instructions are what this station expects. A QA station might say "run the checkout flow on mobile Safari as well as desktop Chrome, and attach a recording of any failure".
Be concrete about evidence. "Verify it works" is not a definition of done; "a screenshot of the success state and the network request that produced it" is.
Triggers
Start work when
Two switches, either or both:
- A card is dropped into this column — dragging a card here, or an earlier station handing one on, starts the work. On by default, and the one that chains stations into a pipeline.
- A card is created in this column — new cards start immediately, including ones an agent writes here.
A comment never starts work; write @ and the agent's name on a card to ask it something. With both off, an agent only works when you ask it to in chat or press Run on a card. That is how the first station of most templates is set up: the analyst is there to talk to, and the cards she writes do not set her off again.
Then move the card
Three selects, each defaulting to Stay in this column:
- When it passes — usually the next station.
- When it fails — usually back to the station that can fix it. The reason is written into the card's discussion on the way.
- When it is blocked or out of retries — the escalation station. Put a person on it.
Each select lists the columns of this workflow and, below them, the stations of every other workflow in the factory. Choosing one of those makes it a route between workflows, and asks what should happen to the card — see Handoffs between workflows.
These are rules, not suggestions to the agent. The agent reports pass, fail or blocked; the system makes the move.
Retry limit
A number from 1 to 10: how many times a single card may come back to this station before it goes to the escalation station instead. It stops a card bouncing between two columns forever. Three is the default: two genuine attempts and one more before a human should look.
Notify by email
An optional address, emailed when a card arrives on this station. Use it on approval and escalation stations — automation nobody is watching is a queue that grows silently.
Handoffs this agent may choose
For work that turns out to belong somewhere else — QA finding a security problem. You list the routes; the agent can hand off along these and no others. Covered in Handoffs between workflows.
Permissions
An agent can always read the card it is working and write cards into its own column. Everything else is off until you switch it on here:
- Open draft pull requests — commits to a new branch in the connected repository and opens a draft pull request. Never merges. Reading the repository needs no permission.
- Use tool connections — tools from the MCP servers you have connected, such as a browser test runner, and only the tools switched on there.
These are grants to this station, not to the agent. The same agent can open pull requests from Development and not from QA.
Runs
Pick a card and press Run to start the worker deliberately, without waiting for a trigger.
Underneath is every run at this station with its status, its result or error, and the steps it took — each tool call included. A live run can be cancelled; a run waiting on a person shows approve and reject.
Removing a worker
Remove worker turns the station back into a plain column. Runs in progress are cancelled; finished runs keep their history, and the cards stay where they are.