While the agent works
While the agent works
Section titled “While the agent works”A long run used to mean a greyed-out message box and a lot of waiting. It doesn’t any more. The conversation stays live while the agent works, and this page covers what the controls do.
Queue the next message, or steer this one
Section titled “Queue the next message, or steer this one”The message box never locks. What happens when you send depends on which key you press:
| Key | What it does |
|---|---|
| Enter | Queues your message as the next turn. A slim Queued · runs next chip appears, with an × to withdraw it. |
| ⌘/Ctrl + Enter | Steers now — the agent picks your message up mid-run and adjusts course. |
The hint under the box says the same thing, so you don’t have to remember.
Use steer for a correction that matters right now — “stop, that’s the wrong repository”. Use queue for the next thing you want done, so you don’t have to sit and watch for the finish. A queued message runs automatically when the current one ends.
If your organization has several runs going at once, a queued message may sit behind them; the indicator tells you how many are ahead. Nothing fails for waiting — although a message that has queued for hours is released with a note asking you to send it again, rather than surprising you by running much later.
Steers are never quietly lost
Section titled “Steers are never quietly lost”If you steer at the very end of a run and it finishes before the agent picks your message up, it isn’t dropped on the floor. It’s carried into one follow-up turn, and the marker on your message says carried. Where it genuinely can’t be carried — you cancelled the run, for instance — it’s marked dropped, with the reason. Either way you can see what happened to it.
Questions you can answer whenever
Section titled “Questions you can answer whenever”Sometimes the agent wants your input but can carry on without it — a preference, a choice between two reasonable options. Those come as a card tagged you can answer anytime — I’m continuing, and the run keeps going.
Answer while it’s still working and it takes your answer into account from that point. Answer after it’s finished and your answer carries into the next turn. Ignore it and nothing breaks.
That is different from a genuine approval, which does stop and wait — see Autonomy & approvals.
When a run reaches its budget
Section titled “When a run reaches its budget”Long runs have a ceiling — a wall-clock budget and a limit on how many steps the agent takes — so a run can never quietly burn your credits forever.
When a run reaches it, it doesn’t just stop. The agent takes one short final turn to tell you where it got to: what’s done, what remains, and what it would do next. Then the conversation shows:
Reached this run’s budget · Continue to keep going
Continue starts a fresh run from that state. Nothing is lost — the plan, the notes and the files are all still there.
When the conversation gets condensed
Section titled “When the conversation gets condensed”Every model has a limit on how much it can hold at once. On a long conversation, Auteryn summarises the earliest part to make room — and tells you it did:
Context condensed — 37 earlier steps summarised · Show
Show opens the summary, so you can check that the thing you care about survived.
You can also do it deliberately. The context ring above the message box has Compact now, with a field asking what must survive? — “keep the API schema and the failing test output verbatim”. The summary is then written to your instruction rather than to a generic one, and the chip afterwards notes that your instructions were applied. Compaction happens at the start of your next message, so nothing is interrupted.
If there’s something an agent should always keep, put it on the agent instead: the Console has a What to keep when the conversation is condensed field per agent, and it applies to every conversation with it.
If something goes wrong mid-run
Section titled “If something goes wrong mid-run”Runs survive the infrastructure underneath them restarting. When a run is picked back up after an interruption:
- Work that already completed is not done again. The results of the steps it had already taken are replayed rather than re-performed, so an email isn’t sent twice and a file isn’t written twice. If the run diverges from what it did before, replay stops there and it carries on live.
- A pending approval survives. If a run was interrupted while waiting for you, the request is still there when you come back, and the conversation makes sense around it rather than showing a half-finished action with no explanation.
What sits above the message box
Section titled “What sits above the message box”Only three things can appear between the conversation and the message box, in this order:
| Surface | When it shows |
|---|---|
| Something waiting on you — an approval, a question, a plan that needs your go-ahead | Only while the run actually needs an answer. One at a time; if several are waiting they stack with a count. |
One line of progress — the current step, 2/4, and, only when they are true, the goal pass (Pass 2 of ≤5) and Preparing computer… |
While a run is working. It expands to the full task list if you want it, and disappears a few seconds after the run finishes. |
| The background bar | Only while something is running in the background. |
Everything else stays in the conversation where it happened. A plan the agent approved for itself is a collapsed one-liner at the turn that wrote it — open it if you want the steps, ignore it if you don’t. A Goal shows as a chip in the progress line, and only becomes a card when a follow-up pass is booked, when it has used more than half its budget, or when it ended without finishing.
If the computer is still being set up when you open a conversation, a small pill says so. If that takes more than three minutes it turns into Still preparing your computer · Retry, and Retry does what it says.
After you press Stop
Section titled “After you press Stop”Stop ends the run that is executing. Anything you typed as a steer that the agent had not picked up yet is dropped and marked as such, so you can see it never landed.
The next message you send starts a new run. It is never treated as a correction to the run you stopped. If you had a message queued behind the stopped run, it stays queued and runs next.