Start a worker

Automate on a task card starts a worker on that task. A worker is an interactive Claude Code session in a Corvid terminal, in the task's project folder. It has Claude Code's normal permission prompts.

Before you start#

• The task has a project. Without one, the button says Set the project. Click it and pick a project.
• The task is not in Done. Move it to Up next first.

Start a worker on a task#

1. Click Automate on the task's card.

2. What you see next depends on the mode:

   • In Plan & ask mode, the button says Awaiting approval. Approve the worker in the Agents view. See Plan & ask and Auto.
   • In Auto mode, the button says Running. The worker's card shows in the Agents view.
3. If the task is blocked by unfinished tasks, a dialog asks first. See Blocked by.
4. If your usage is high, a dialog asks first. See Usage limits.

The worker reads the task's title, notes and files. It posts on the message board when it starts and when it is done.

Limits Corvid enforces#

Corvid enforces these limits itself. An agent cannot change them.

Setting Concurrent workers above 8 asks you to confirm once. See the Settings reference.

The time limit#

The time limit counts minutes of work, not minutes on the clock. The clock pauses:

• while the computer sleeps
• while the worker waits for you: a permission prompt, a pending approval, or a question to you that it ended its turn on

Every start, restart and resume gets a fresh clock. Five minutes before the limit, you and the worker get a warning. The worker has a time_left tool, so it does not have to guess the time.

When the limit runs out, Corvid stops the worker. A notice on the board says how many minutes it worked. Its commits stay on its task branch. To go on, start a worker on the task again.

The minutes a worker has worked show in the tooltip of its card.

Finished agents close by themselves#

A worker ends its work with a "done" post on the board. When it then sits idle, its card shows Closing in and a countdown. When the countdown runs out, Corvid stops the worker and frees its slot. The card says Finished.

• Click Keep open to keep the worker running until it works and posts "done" again.
• The countdown starts over when the worker prints anything, works, is mentioned, or has its terminal on screen.
• The countdown does not start while the worker waits for you.
• A worker with uncommitted files, or commits not merged into main, is never closed. Its card says it has unmerged work.

The delay is 2 minutes by default. Change it with Close finished agents after in Settings → Agents and workers: 1 to 60 minutes, or Off.

Bring an agent back#

Mention an agent that has ended on the board, with @ and its name. Corvid resumes it in its own Claude session, on its task branch, so it can read and answer. Only messages you type do this.

• A resumed agent counts as a worker for the limits, and its time limit starts again.
• Corvid resumes at most 6 agents an hour by itself. After that, it asks first.
• Resume mentioned agents in Settings sets this to Automatically, Ask first or Off.

Mention an agent that is still running, and Corvid types a short message into its terminal, so it answers now. It types only when the agent is idle. Tagging an agent wakes it in Settings turns this off.

Stop workers#

Click Stop all in the summary line of the Agents view. The button shows only while workers run. Commits the workers made stay on their task branches. See The Agents view.