Autorun

Autorun starts a new task or sends an instruction to an existing task at a specified time, interval, or cron schedule. It does not coordinate several agents inside one run. It is an independent way to start or continue the same kind of work when it is needed.

Use Hooks to react to events such as task completion or an Ask, rather than a time or interval.

Requirements

Autorun requires an AGI Labo membership. Cockpit verifies membership not only at the Desktop entry point, but also in the PWA, cockpit autorun, the local API, and each scheduled execution. Authentication that needs renewal, an inactive membership, and a temporary verification failure produce distinct errors during an operation. Cockpit rejects an operation safely when a network or API failure prevents verification, but it does not cache that failure and block a member after service recovers. Existing members can use the CLI while signed in to AGI Labo in Desktop.

Schedules run inside the AGI Cockpit app process. Closing every window on macOS does not stop the scheduler. The operating-system scheduler does not launch Cockpit after the process exits, but Cockpit performs one catch-up per Autorun when it restarts or resumes within 24 hours of an unhandled scheduled time. Older occurrences are recorded as missed instead of silently starting stale work.

Schedule types

In the cron weekday field, 0 means Sunday. For example, 0 9 * * 1-5 runs at 9:00 a.m. on weekdays.

Create an Autorun

1. Open the app menu in the lower-left corner of Desktop, then select Autorun tasks.
2. Select New Autorun.
3. Enter a name and the instruction to send when the task starts.
4. Under Execution target, choose Create a new task or Send to an existing task.
5. For a new task, choose the working directory and agent, then review any shown UI mode, account, model, reasoning effort, service tier, system prompt, and approval mode. Cockpit uses the Master's working location for Master Agent runs, and Auto is the default for agents with account selection.
6. For an existing task, choose Target task. Cockpit creates no new runtime; it uses the target's existing agent, conversation, working directory, and runtime settings.
7. Choose Once, Interval, or Cron and configure the timing.
8. Save the Autorun and confirm that the list shows its next run time.

In Desktop and the PWA, open Target task and search by task name, project, instruction text, or task ID. Review the name, status, project, and shortened ID in the results, then select the target. A unique ID prefix resolves to the full task ID; when several tasks match, select the intended result. Arrow keys move through the results, Enter selects, and Escape closes the picker.

Wait for the task list to finish loading before checking an unknown ID. If Cockpit offers Use the ID … anyway, review the warning before selecting it. Saving an unknown ID is allowed, but the Autorun is disabled if the task is missing when execution starts.

When creating or updating an Autorun in the PWA, the dialog stays open and disables its controls until a save result arrives. A failed save preserves the entered values, keeps the dialog open, and displays the reason returned by the server. A membership failure links to plan selection, while a disconnected result says that nothing was saved. A timeout means the result is unknown: close the dialog and inspect the list before retrying so you do not create a duplicate.

An Autorun that creates a new task keeps its runtime settings from the time it is saved as a snapshot. Changing global settings later does not silently change its model, reasoning effort, service tier, system prompt, approval mode, account-selection method, or UI mode. If a pinned model or account becomes unavailable, Cockpit does not substitute another setting. It disables that Autorun and marks it as needing attention.

A new-task Autorun also keeps its Browser Identity assignment and passes it to every task the schedule creates. An Autorun without an explicit assignment uses the Default Identity. In v4.43.0, assign or change an Autorun's Identity from the CLI. Create, rename, recolor, clear, or remove the Identity itself from Browser Identity in the lower-left app menu or from the CLI. An existing-task Autorun does not change the Browser Identity already assigned to its target.

For a new task, Desktop and the PWA show only settings supported by the selected agent and UI mode. The CLI and API use the same capability data and reject unsupported combinations with an explicit error before saving.

Before saving a new-task Autorun, Cockpit checks the selected model and reasoning effort against the available catalog. If a value is unavailable, saving is blocked and the affected setting is shown; select a supported value or reload the list after updating the agent CLI. A failed lookup can show built-in fallback candidates, so inspect the list’s source and retrieval status. Unlisted Codex models follow the exception described below; reasoning values are still validated. These checks do not change the runtime of an existing-task Autorun.

For Codex Native UI, selecting a pinned account loads the latest model catalog available to that account and lets the Autorun store its supported reasoning level and service tier. While editing, Auto uses the default account's catalog; at execution, Cockpit revalidates the stored values against the catalog of the account that Auto selects. Saving waits while discovery is in progress, and a failed discovery falls back to the built-in candidates. Cockpit Agent likewise reflects models from the connected OpenRouter, OpenCode Go, OpenCode Zen, or LM Studio provider and exposes only the reasoning levels advertised for the selected model.

Codex can retain a saved model as not in list when it is absent from the retrieved catalog. If capabilities are unknown, the picker offers low, medium, high, xhigh, max, and ultra. Absence alone does not block saving; Codex decides actual availability at runtime. See Agents and models to check a custom catalog.

Cursor Autoruns can use Native UI or Terminal. Native UI exposes Cursor's available model, model-specific reasoning effort, approval mode, and account profiles in Desktop, the PWA, and CLI-backed creation. Cursor does not expose service-tier or system-prompt settings.

Qoder Autoruns can also use Native UI or Terminal. Native UI exposes Qoder's available model, model-specific reasoning effort, system prompt, approval mode, and account profiles in Desktop, the PWA, and CLI-backed creation. Qoder does not expose service-tier settings.

Claude, Codex, Grok Build, Antigravity, Cursor, and Qoder Autoruns store either Auto or a pinned account. Auto is the default for new Autoruns. It is stored as a selection method rather than as one concrete profile, so every execution chooses again from the usage state of signed-in accounts. If the created task reaches a usage limit or returns a plan-restriction response, Cockpit also switches to another available account and continues processing. Use --account <name|id|default> to pin an account from the CLI and --account auto to select Auto.

Review a run

When a new-task Autorun fires, it creates a task in the same form as any regular task. Select it from the task list, then use task details to review its conversation, confirmation requests, errors, and results. The new task keeps a reference to the Autorun ID that created it.

A new-task Autorun can also complete the task it created once that run finishes. The option is off by default; turn it on per Autorun in the Autorun dialog, in the PWA, or with --complete-on-finish true. Cockpit treats a turn_complete with no Ask, no permission prompt, and no usage limit as the end of the run, and does not depend on the agent running cockpit task complete. A task stopped by an Ask, a permission prompt, a usage limit, a process error, or any other state that waits for user input stays open. Existing-task Autoruns never complete their target, and Terminal Autoruns do not offer the option because a terminal task never reports turn_complete.

An existing-task Autorun sends its text as that task's next instruction. If the target can accept input, delivery happens immediately. If it is running or stopped by a usage limit, Cockpit stores the instruction in a persistent main-process queue until the task becomes ready. The queue has no timeout, and completed or errored targets are resumed automatically when possible. Delivery fails, disables the Autorun, and marks it as needing attention if the target is deleted, cannot be resumed, or is a Windows Terminal task whose foreground process cannot be verified safely.

The Autorun manager supports:

• enable and disable
• edit
• delete
• Run Now
• previous and next run times
• the latest run status and the created or target task ID

Every scheduled occurrence records one of created, queued, delivered, skipped-membership, skipped-lock, skipped-period, missed, or failed. queued means an existing-task instruction is waiting; delivered means it was sent. Authentication or membership skips, missed runs, and startup or delivery failures are marked as needing attention in Desktop and the PWA and produce a desktop notification. Overlap and period-dedup skips remain informational. cockpit autorun list and cockpit autorun get expose the same lastRunStatus, lastRunScheduledAt, lastRunDetail, and lastTaskId fields, plus executionTarget and targetTaskId. nextRunAt is the next attempt time, while nextRunScheduledAt preserves the original occurrence during retries. Cockpit also appends one JSON line per occurrence to data/cockpit/logs/autorun-executions.jsonl under the AGI Tools data directory. A queued occurrence receives a later delivery or failure record. Repeated retries with the same outcome are recorded once; the log rotates at 5 MB and keeps two archives.

For a new-task Autorun, Cockpit skips a scheduled run only while another launch is still starting. The lock covers the launch itself and never waits for the created task to finish, so the next occurrence creates a new task even while the previous one is still running or waiting for confirmation. A 24-hour expiry is the final safeguard against a launch that never settles. Existing-task delivery releases the lock the same way once the occurrence is queued or delivered; pending instructions are processed in order for each target task.

Working directory and failures

For a new task, the Desktop form requires a directory unless the Autorun runs as the Master Agent. If the CLI omits the directory, Cockpit creates an agi-cockpit directory inside the operating-system temporary folder. Existing-task delivery uses the target's working directory and does not take a separate Autorun directory.

If the agent fails to start, the task remains in the task list with an Error state. A one-time Autorun still becomes disabled; an interval or cron Autorun remains enabled and calculates its next run. Open the created task details to inspect the cause.

If Cockpit confirms that the account is a guest or has an inactive membership when a run is due, a one-time Autorun does not start and becomes disabled. An interval or cron Autorun remains enabled and checks again at its next scheduled time. If the sign-in session needs renewal or a temporary network or API failure prevents membership verification, Cockpit does not start the run and checks again one minute later. It keeps a one-time Autorun enabled and persists the retry time, so authentication renewal or a temporary outage does not consume its only run.

Create from the CLI

cockpit autorun create \
  --name "Weekday progress review" \
  --instruction "Review the unfinished work in this project and summarize it in priority order." \
  --directory /path/to/project \
  --agent-type codex \
  --account auto \
  --ui-mode visual \
  --model gpt-5.4 \
  --effort high \
  --service-tier fast \
  --approval-mode accept-edits \
  --browser-identity work \
  --type cron \
  --expression "0 9 * * 1-5"

cockpit autorun list
cockpit autorun get <id>
cockpit autorun run <id>
cockpit autorun toggle <id>
cockpit autorun update <id> --browser-identity default

To send to an existing task, specify the execution target and task ID:

cockpit autorun create \
  --name "Continue review" \
  --instruction "Check progress and continue with the next step." \
  --execution-target existing-task \
  --target-task <task-id> \
  --type interval \
  --minutes 30

Related pages

• Task list
• Task details
• Ask