Steering a run
Talk to the agent while it is writing, get the remaining phases revised, and know what happens to the conversation when the run ends.
The agent is behind a flag and is not switched on in the studio yet. See the agent's desk.
A run does not go quiet while it writes. The prompt field stays open, and what you type there goes into the run's thread. At the next phase boundary the agent reads it.
That is steering: not an interrupt, an input queue.
Say something mid-run
- Keep the agent thread open while the run is writing.
- Type into the prompt field and press Enter, the same as any other agent message.
- Your message appears in the thread straight away.
What happens next depends on the message. A question ("how's it going?", "what does the courier want again?") is answered in the thread while the run keeps going. A correction ("stop using the warehouse, use the docks") is carried into the next re-plan.
At each phase boundary the agent is asked whether to keep the remaining phases or replace them, and anything you have said since it last looked is handed over as "The author said meanwhile: …". It either revises the remaining phases, in which case a new plan appears in the thread, or it keeps them and answers you.
Steering never aborts anything. The phase that is running finishes. If you want the run to stop, press Stop, which also lands at the next phase boundary. Steering is strictly weaker: it informs, it does not cancel.
What steering cannot do
It cannot reach the last phase. There is no boundary after the final phase, so a message that arrives while the last phase is executing is not consumed by this run. It is not lost either: it stays in the thread and becomes the opening of the next exchange in that session, on the same review branch.
It cannot start a second run. A prompt that reads as a new task while a run is writing is refused with "A run is already in progress on this project." Wait for the run to reach awaiting review, then send it as a follow-up, which appends to the same branch and the same ledger.
It cannot undo written commands. Phases already executed stay in the ledger. To remove them, review the draft and use Rewind to here, described in runs and review.
Steering that lands well
- Say it as a constraint, not a rewrite instruction: "keep Mira out of the docks scenes" beats "redo phase 3".
- Send it early. A correction sent during phase 2 of 5 has three boundaries left to act on. The same words during the last phase wait for the next session.
- One thought per message. The re-plan gets your messages as a list, and a paragraph of mixed requests tends to survive only in part.
When the run ends
The conversation outlives the run. The thread keeps every exchange, the plan, its revisions, and the ledger it produced.
While the draft awaits review, the session is open. A new prompt continues it: the agent runs again on the same review branch, appending its commands to the same ledger. This is where "now do the same for chapter four" belongs.
Once you merge or discard, the session closes. The next prompt is refused with "This session is closed, its draft was merged or discarded. Start a new one." The studio opens a fresh conversation with the same prompt, forked from current main, so you do not have to retype it.
Chat that never became a task lives in its own thread with no branch behind it, and it is restored when you reopen the project. The agent reads that history when you follow up, which is why "let's do 2" means something after it lists three options.
Where to look when nothing seems to happen
- The chip reads
writing · batch N/M. It has not reached a boundary yet, so your message is queued. - The chip reads "stopping after this phase…". Stop was pressed, and only the current phase remains.
- The chip reads "awaiting review". The run is over. Your message is a follow-up prompt now, and it will start a new run on the same branch.