Autonomous missions
By default an agent runs a single turn and comes to rest. Flip Run autonomously on when you start it (off by default) and the task becomes a durable objective: the agent pursues it on its own, continuing between turns and pulling you in only when it finishes or gets stuck.
- Objective, not a one-shot prompt. A mission strip at the top of the timeline keeps the standing objective in view with a
step N/maxcounter. - The goal loop. Each time the agent goes idle, Precursor reads the tail of its reply for a control directive and decides what happens next — hand back, stop, or keep going. No repost fires between steps, so a multi-step mission lands as one result in the linked topic/chat at the end.
- Self-reported progress. A
PROGRESS:directive renders a progress bar with a percentage and short label on the card and mission strip. - Blocked, not silently looping. If the agent needs a decision it emits
NEED_INPUT:and moves to a Needs input state, surfaced as an amber callout with an Answer button; your reply resumes the mission. It also self-blocks if it stalls or spends its step budget, so it can't spin forever. - Step budget. A per-agent step budget (default 12, 1-100) caps how many times it may continue before handing back — the primary guardrail.
The autonomy protocol
An autonomous agent is told to end a reply with a lifecycle directive; the goal loop parses them (case-insensitive, last match wins).
| Directive | Effect |
|---|---|
OBJECTIVE_COMPLETE: <summary> | Mission done — status Completed, summary saved, progress forced to 100%, single repost to the linked surface. |
NEED_INPUT: <question> | Agent is blocked — status Needs input, the question surfaced for you; your next message unblocks it. |
PROGRESS: <0-100> | <label> | Optional heartbeat — updates the progress bar and resets the stall counter. |
ARTIFACT: <title> | <content> | Optional, repeatable — publishes a named output to the shared blackboard. Doesn't change the run's lifecycle. |
All four marker lines are stripped from the rendered transcript and re-surfaced as structured UI, so the prose stays clean. If none appears and the budget/stall guards allow it, the agent continues to the next step. Plain (non-autonomous) agents ignore the protocol entirely and rest at Idle.
Experimental
Autonomous missions are new and best kept on a step budget with permission guards in place. Watch a mission's first few runs before trusting it unattended.