Autonomy

What it does

• Three modes

  One turn at a time, a prompt written for you, or a prompt that sends itself.

• A policy, not a guess

  Six answers set what may be picked up, what counts as done, and how the loop reacts.

• Failures steer the loop

  A failed check or a blocked task goes into the next prompt, so the work moves toward the goal instead of repeating.

• A real stop condition

  The loop ends on the completion marker, when the tasks are done, after eight automatic steps, or on a repeated failure.

• You can still stop it

  Escape twice ends the turn, and the transcript keeps a quiet stopped rule where it happened.

• Fixed or written

  The next prompt can come from the six scenario prompts in the policy or from a smaller writer model.

Pick a mode

The robot in the prompt bar opens the three modes. Pick one, and the same turn runs again with a prompt written for the next step.

• Agent loop

  One turn at a time. Pip answers, files nothing extra, and stops.

• Semi-autonomous

  Pip writes the next prompt. You read it and press send.

• Self prompting

  Pip writes the prompt and sends it itself, until the policy says stop.

Policy

A policy is six answers. They compile into one system contract that the agent carries for the whole run, plus one fixed prompt for each point in the loop: continue, error, question, complete, checkpoint and stuck.

• What may start

  The open project, the next task only, or the tasks you have marked ready.

• Where it stops

  Every task closed, every check passing, or a product that runs end to end. A failed check that crosses your line pauses the loop.

• When it gets stuck

  Ask and wait, replan once and retry, or close the run with what it learned.