Decide an approval, and close the loop with feedback

Two task types exist for the two ends of a decision: the one that asks you before something happens, and the one that reports back afterwards.

Approval tasks

An approval task is the platform asking a question it is not allowed to answer. It carries the proposed action, and the work that raised it waits rather than guessing.

This page uses LNCH-5, the Summitline waitlist email waiting on Maya.

Reading one

The task’s Overview shows what is being proposed and, where there is one, the reason. The kinds of action you will see:

Kind	Example
publish	Send the waitlist email to the beta list
tool call	Use a tool the agent’s autonomy level will not let it use alone
budget override	Spend past a run’s cap
state transition	Start a run for an assigned task

Deciding one

1. From the task, select Open in Approvals (or Open run gate, when the decision belongs to a paused run).

2. In the queue, read the request, then Approve or Deny. A denial takes a reason.

3. The work that was waiting continues, or does not, on the answer.

Approve and deny live in the approvals queue and on the run gate, not on the task page. One decision, one place to make it, whether it came from a content pipeline or a paused run.

The four outcomes

pending while it waits, then approved, rejected, or expired. For a content task, approved sends the draft to scheduling, rejected sends it back to drafting, and expired kills it with that reason recorded.

You do not create approvals

Approval tasks are raised by the platform when an agent hits something its autonomy level will not let it do alone, so there is no approval entry in the create menu. What you control is the ladder: raise a class and fewer approvals appear, lower it and more do.

Feedback tasks

A feedback task is the other end of the loop. Something shipped, and the question is what happened.

It carries:

• A subject, the thing being measured, usually a published post or a completed run.
• Metric sources, where the numbers come from.
• Collection windows, the times at which to look. Several windows means a day-one number and a day-seven number rather than one snapshot.

The lifecycle

State	Means
awaiting window	Waiting for the next collection time
collecting	Reading the metric sources
collected	Numbers are in
learned	Read and closed, with the conclusion recorded

Collection retries on its own when a source is briefly unavailable, and a window that keeps failing records the error rather than a zero. A window with no data says so.

Reaching learned is what closes the parent content task as done. Results come back as work, not as a dashboard nobody opens.

Next steps

• The autonomy ladder: what turns an action into an approval.
• Run a content task: the pipeline that spawns both of these.
• Act on notifications: how a decision reaches you.