Autonomous board reviews

The PM agent can also work without being asked. This is off until you switch it on, and it is deliberately narrow.

Scheduled reviews

Under Settings → PM agent → When it acts on its own:

• Review the board on a schedule — the switch.
• First review at — the hour of day it starts.
• How often — the interval in hours, up to once a day.
• Timezone — the project’s own timezone, so a review at 9am is 9am for the team, not for the server.

Each scheduled slot is claimed atomically, so a review runs at most once per slot no matter how many app instances are running.

Running a review now

Run a review now, in the same section, starts one review of the board as it is, without waiting for the schedule, and whether or not the schedule is switched on. It is the quickest way to see what a review will say before you rely on it. Its report lands in the PM chat marked Scheduled review, like any other.

It takes the same path as a scheduled one, so it is refused, and says why, when the PM agent is off or locked for the project, when the instance has no model credentials configured, when the day’s turn or token cap is reached, or while another PM turn is running on the project. It uses a turn from the daily cap. It does not use up the next scheduled slot: the schedule still runs as set. Only a project owner can start one, from a browser session; an API token cannot.

What a review actually sees

The agent does not wander the board unaided. The server computes a digest first and hands it over: how many tasks are open and how they sit across the columns, plus three findings —

• tasks missing acceptance criteria or a description,
• tasks stuck in a column longer than they should be,
• duplicate titles.

During a board review, scheduled or run now, the agent runs with change_status, create_task and assign_task withheld. It can comment and it can improve a task’s content — it cannot move your board, invent work, or hand anything to a person or a machine.

Handling human-review tasks

Review tasks that land in “Needs human review” is the second trigger. A column can be marked to trigger a PM review, and tasks arriving there are queued for the agent to look at — so an agent-written change that needs a decision gets a first-pass summary before a person opens it.

This trigger withholds a different pair: change_status and assign_task. It may therefore create a task, which is the point — a review that finds follow-up work can write it down. What it still cannot do is move the task on or put somebody’s name against it.

Cost

Autonomous turns count against the same turns per day limit as your chat, and against the same tokens per day ceiling. When either is reached the agent stops until the next day — it does not queue up and spend tomorrow’s.

A review is a full turn, so it is up to fifteen model calls on its own. If you switch reviews on for a busy board, watch the day’s token figure in Settings → PM agent for a few days before deciding the limits are right — see what a turn costs.