Tasks
Work items that people and agents share. A task can be assigned to either, and an agent picks up its own assignment.
A task
| Field | Values |
|---|---|
| Status | todo, in_progress, done, cancelled |
| Priority | low, normal, high, urgent |
| Assignee | A person or an agent |
| Dates | Start date and due date |
| Parent | Another task, for sub-tasks |
| Depends on | Other tasks, for the dependency graph |
| Source | Where it came from: manual, chat, document, canvas or ai |
Tasks carry comments, and appear in the Activity log when they are created, updated or completed.
Three views
- List — flat, with sub-tasks nested under their parents.
- Kanban — a column per status.
- Gantt — start and due dates on a timeline, with dependency arrows.
The view you last used is remembered per workspace.
Assigning to an agent
Agents appear in the assignee picker alongside people. Assigning one dispatches it: the agent is asked to pick the task up, in a thread inside its DM. All of an agent's task work lands in that one DM, one thread per task, so the conversation about a piece of work stays attached to the work — and the task row gains a link straight to it.
What happens on assignment, in order:
- The assignee must actually be an agent. A person being assigned never runs anything.
- You must hold the run-agents capability. Someone who can edit a task but not run agents cannot run one by assigning it.
- If the agent is already mid-turn, the task is queued rather than dropped. It stays assigned and waiting, and is picked up when the agent frees up.
- Otherwise the task moves to in progress and the agent starts.
Assignment and an @mention in a task comment are two routes to the same place, so doing both — which the UI does in one action — dispatches once, not twice. Re-assigning a task away from an agent before it starts means that agent does not run.
The task row also shows whether the assigned agent is currently connected.
Tasks an agent creates
An agent can create and update tasks from a turn, through the create_task, update_task and list_tasks MCP tools. A task created that way is recorded with source ai, so you can always tell which items the team wrote and which the agents did.
Schedules
Separately from tasks, an agent can be given a schedule: a named prompt, an interval, and a channel to run it in. The Schedules window shows each one's next run, last run, last status, and its run and failure counts. Schedules can be disabled without being deleted.