Agent Background Work

  • Desktop, mobile and web
  • All plans

What it does

An agent can start sub-agents and background commands, then go back to its prompt. It looks finished while that work keeps running, and closing the agent stops it. AgentsRoom shows what is still running behind the agent, and warns you before a close that would stop it.

Where to find it

  • Agent row: a small pill with a layers icon and a count, next to the agent's name. Hover it for the summary and the names.
  • Above the composer: an In background chip with the summary, for as long as something runs. Click it to open the Activity tab, filtered on this agent.
  • Right panel, Activity tab: an In background list at the top, with each task's name, a green Live mark, the agent, its type and when it started. Click a task to open its detail. The tab shows a green dot while something runs.
  • Wherever the app spells out an agent's status (session details, the terminal preview from the backlog, the command palette): In background replaces Done or Inactive while a background command runs. An agent waiting on a background sub-agent or workflow stays under Working until it ends.
  • An agent running on another of your computers (Remote Fleet): the same pill next to its name, In background instead of Done on its row and on its console card in a team run, and the same chip above its composer. The tooltip gives the count ("1 command"); the task names stay on the computer that runs it.

How to use it

Nothing to turn on. When you close an agent that has something running, a dialog lists the tasks. Keep it running cancels the close (also Esc). Close anyway closes the agent and stops them.

The pill and the chip disappear by themselves when the work ends.

The detail of a task opens in a modal, read-only and live, from a row of the Activity tab. A command shows its full command line and its output. A sub-agent shows its type, the prompt it received and its conversation, with its tool calls. The view follows the end like a terminal until you scroll up, the copy button takes the text, and Esc closes it. It stays open when the task ends, marked Ended, so you can read its last lines. A long log shows its end only, and says so.

Settings

No setting. The Activity tab can be hidden with uiElements.rightPanel.tabActivity (see Agent Live Activity); the pill, the chip and the warning stay.

Agent tools (MCP)

agents_close refuses to close an agent with background work unless called with force: true. agents_delete skips it, with the reason. Automatic pauses and batch restarts after a CLI update leave it alone.

Providers

  • Claude Code: sub-agents and background commands, including Monitor. The detail of a task shows its output or its conversation.
  • Grok Build: sub-agents and background commands. The detail of a task shows its command, and its output only when the CLI reports where it writes it.
  • Codex, Antigravity, Cursor, OpenCode, Copilot and the others: nothing is shown yet, their CLI gives no reliable signal for this.

An agent started before this version must be relaunched: the CLI reads its hooks when it starts.

Mobile

Yes. Same pill on the agent's row, an In background section on the agent card (tap the avatar or name in its console) with the same Live mark, the same detail when you tap a task in that section (secrets masked, a shorter end of the log, the sub-agent's prompt and its last messages), and the same warning before closing the session from the actions sheet, the manage sheet or a swipe. In the Remote screen, an agent of another computer shows the same pill and In background instead of Done.

Limits

  • Only what the CLI declares as background work. A process the agent detached by hand (nohup, & inside a command) is not listed.
  • An agent on a remote machine over SSH shows nothing: its hooks cannot reach this computer.
  • An agent on another of your computers needs AgentsRoom up to date on the computer that runs it, since that computer is the one reporting. Stopping it from here does not show the warning.
  • Close All Sessions, closing a project, quitting the app and the automatic close of a finished ticket do not show this warning.
  • Kept in memory on the computer that runs the agent, gone with the console.
  • When the agent's .claude/settings.local.json disappears while its CLI runs (an agent that removes its own worktree on Windows, a file deleted by hand), the CLI stops reporting anything to AgentsRoom. The list of that agent is then emptied, even for a task still running, rather than showing a task as Live for good. A new console of the agent reports again.
  • The detail of a task reads the file the CLI writes. When a CLI writes none, or a task started before this version, the modal says the CLI does not share the output. Nothing is kept once the CLI removes that file.

Common questions

Can I see what a background command prints, or what a sub-agent is doing? Yes. Open the Activity tab and click the task in the In background list: the modal shows the output of a command, or the prompt and the conversation of a sub-agent, live.

My agent says it is done but the pill still shows a number. Something it started is still running. Hover the pill or open the Activity tab to see what. It disappears when that work ends.

Does closing the agent stop its background commands? Yes, they run under the agent's CLI and stop with it. That is why AgentsRoom asks first.

My agent says In background instead of Done, is it working? Its turn is over and it can take a new prompt, which is why it stays under Available and not under Working. A command it started is still running behind it. A dev server left running would otherwise keep it marked as working for days.

Why does my agent stay under Working when its terminal shows it waiting for a background agent? Because the work is not over: a sub-agent or a workflow is still running, and the agent wakes up when it ends. It goes back to To review when that last turn ends. A background command alone does not do this. You can still type to the agent, and its messages from other agents still reach it.

The pill shows fewer commands than my terminal. Commands started by a sub-agent count as that sub-agent while it runs: open it from the Activity tab to read them in its conversation.

Why does my Codex agent show nothing? Codex does not expose a reliable signal for this yet. Claude Code and Grok Build do.

A finished command stays Live after my agent removed its own worktree. Fixed on 2026-11-13 (reported by Андрій): on Windows, removing the worktree an agent runs in empties its folder, settings included, and the CLI stopped telling AgentsRoom that the command had ended. The list now empties itself at that moment. Start a new console of the agent for the fix to apply.

Can I stop only one background task from AgentsRoom? Not yet. Ask the agent to stop it, it has the tool for that.