# Background runs and schedules

> Start work from @handle, schedule repeated tasks, and understand unattended permissions.

An agent can run without an AI Chat window. Type its `@handle` followed by a task in root search, select it and run it. The palette closes, an Agents menu bar indicator shows activity, and the result is saved as an ordinary chat.

```text
@docs Summarize the changes in this project's README and list anything that needs updating.
```

Use a handle configured for one of your agents. The task runs in that agent's folder with its instructions, model and permissions.

## While and after it runs

The Agents menu lists active work and can also include live coding-agent sessions. Open the run to follow it in AI Chat. A completion HUD offers **Open**; Notification Center banners are an optional per-agent setting and name the run rather than displaying its answer.

A background native agent uses its effort's time and tool budget. An ACP agent uses its own runtime, with Redcast's unattended delivery and permission handling.

## Create a schedule

Open the agent's scheduling controls and choose a prompt, frequency and missed-run policy. Schedules can run daily, on working days or every few hours. A schedule can also append its answer to a selected note.

Prefer a self-contained prompt: identify the files or sources, the expected report and what counts as success. There may be nobody available to answer questions when it runs.

## Sleep, quit and time zones

Schedules run **only while Redcast is running**. There is no launchd service that keeps executing after you quit. Times follow the Mac's local wall clock and adjust after clock or time-zone changes.

After launch or wake, a due run waits briefly for the network. **Run once when back** collapses missed occurrences into one catch-up run. **Skip it** discards missed occurrences. Creating, retiming or re-enabling a schedule starts its timing from that point; it does not replay old occurrences.

If the same agent is busy, a due run waits for its existing run to finish.

## Permission questions in the background

The agent's **In the background** choice controls an operation that would normally ask:

- **Ask me** may show a dialog when the Mac has been used recently. It does not activate Redcast over your work, and Return defaults to declining.
- **Decline** refuses operations that would need a question.
- When nobody is present, an asking operation is declined rather than left waiting indefinitely.

**Always Allow** and **Never Allow** rules still apply. Only MCP servers already trusted as **Always** are supplied to unattended runs. Screen tools and delegated agents cannot silently capture the desktop.

## Daily spend cap

A daily cap counts **reported costs of scheduled runs** for the local day. New scheduled runs stop starting after the cap is reached; reported costs can also stop an active capped run. Costs already reported count even if the run is cancelled or fails.

A route that reports no cost contributes no monetary amount. The cap is therefore not a guaranteed billing limit for every provider. Check [AI Usage](/docs/ai/usage) and the provider's own limits.

**Run Now** and manually typed `@handle` tasks are explicit starts: they are not held or counted by the schedule cap.

## Backups

Full backup can carry agent and schedule configuration, but imported schedules arrive **paused**. Review the folder, command, model, permissions and output destination before enabling them on another Mac. Existing consented schedules on that Mac are treated separately during import.