Background agent runs

When you send a message to the Compass in-app agent, the work happens on the server, not in your browser. You can close the tab, switch conversations, put the laptop to sleep, or lose your connection — the agent keeps going, and the reply is waiting in the conversation when you come back.

This also means a turn is no longer limited to what fits in the few minutes a browser will hold a request open. Ask for something substantial.

While a run is going

The conversation shows the agent's progress as it happens: the text it is writing and the Compass tools it is using, in order.

Underneath the transcript you will see a note — This runs on the server, you can close this tab and pick the reply up here later. That note is the signal that the turn is detached. While it is showing:

• The Send button becomes a Stop button. Sending another message has to wait: a conversation runs one turn at a time.
• Closing the tab, navigating away, or reloading does not cancel anything.
• You can watch the same run from more than one place at once — a second tab, or your phone — and both views show the same transcript.

Coming back to a run

Reopen the conversation and Compass reattaches automatically. If the run is still going, the live view resumes, including everything that happened while you were away. If it finished, the reply and the tool steps it took are simply part of the conversation.

Nothing is lost in between: the agent records each step as it happens rather than only at the end, so a run you never watched has the same transcript as one you watched throughout.

Links to Compass items appear as cards

When the agent links to something in your workspace — an opportunity, task, doc, key result and so on — the link shows as a card with the item's current title, status and a line of context, rather than a bare URL. Click the card to open the item.

Cards are live: they refresh when you return to the tab, so a status changed elsewhere shows up without re-asking the agent. A link to another site, to a page rather than one item, or to something you don't have access to stays an ordinary link.

Sharing a conversation

A conversation starts private to the person who began it. The owner, or anyone already in the conversation, can invite other members of the workspace from the people panel; there is no limit on how many. Invited people see the whole conversation, can send messages, and can watch a run as it happens.

• Each message shows who sent it, and the agent is told who said what, so it can tell two colleagues apart.
• Usage from a turn is counted against the person who sent it.
• Still one turn at a time: if someone else's run is going, you cannot send until it finishes (you are told the agent is busy).
• Only the person who started a run, or the conversation owner, can Stop it.
• The owner can remove anyone; anyone else can leave. Someone who is removed, or who leaves the workspace, loses access straight away.
• Other people's messages and replies appear within a few seconds. There is no live push, so a brief delay is expected.
• Conversations the agent starts from a hand-off or interview stay private to the person who began them and cannot be shared.

Stopping a run

Stop ends the run on the server and releases its resources. The conversation records that it was canceled.

Stopping is not an undo. Anything the agent already did — a task it created, a comment it added — stays done. Check the conversation's tool steps to see what that was, and look at Agent activity for the full record.

If the agent finishes in the moment between your clicking Stop and the request arriving, the result is kept rather than thrown away. The conversation shows what actually happened.

Time limits, and runs that end on their own

Every run gets a time budget — 20 minutes by default, and never more than one hour. A run that reaches its budget stops, and the conversation says so. Whatever the agent produced before that point is preserved.

A run can also end on its own without a reply, if the machine it was running on fails. Compass notices at its next daily sweep and marks the run interrupted rather than leaving the conversation looking busy forever. Your transcript up to that point is kept, and you can send the message again.

Daily usage limits are unchanged, and they apply when a run starts. A long run does not consume more of your daily allowance than a short one.

When a turn fails

A turn that fails while it is running inside the browser request now reports why. If the sandbox wrote anything to its error output before the failure, the first part of that output is appended to the error message, so a failure reads as a cause rather than only an exit code. Error output from a turn that succeeds is ignored. Output is capped, so a very noisy failure shows its beginning, not its end. Runs that execute in the background do not include sandbox error output.

A blank or whitespace-only ANTHROPIC_API_KEY is treated the same as a missing one: the turn is rejected up front with a configuration error, before a sandbox is started.

Deployment

Administrators must apply and verify migration 069_background_agent_runs before this behavior is available. Until it is applied, the agent keeps working exactly as it did before — turns run inside the browser request, with the previous few-minute ceiling — so deploying ahead of the migration is safe and changes nothing for users.

Two settings matter for a deployed environment:

• CRON_SECRET must be set. Compass runs a scheduled job once a day (the Vercel plan only allows daily crons) (/api/cron/agent-run-sweeper) that ends runs which have exceeded their budget or whose sandbox has stopped reporting. Without the secret the job refuses to run, and interrupted runs stay visible as in-progress until it is configured.
• The scheduled job is declared in vercel.json. A deployment that drops that file keeps working, but nothing cleans up after a failed run.