Autopilot

Autopilot lets your agents execute one-click recommendations directly against your connected integrations - without you in the loop. It is gated to Growth and Studio plans, off by default, and bounded by hard safety rails.

What Autopilot can ship

Autopilot only runs for action classes that meet three criteria: deterministic execution, automatic rollback, and bounded blast radius. The current allow-list:

Recommendations from other classes (creative rotations, engineering bug surfacing, pricing changes above 15%, screenshot rotations) always require manual approval - Autopilot does not touch them.

Enabling Autopilot per action class

Autopilot is configured per app, per action class. Open Settings → Autopilot for each connected app and toggle the classes you trust. Defaults are conservative - only paywall variant rollouts on existing experiments are on out of the box.

Each toggle has a tier filter: you can require, for example, that Autopilot only fires on High confidence recommendations, or only when the predicted MRR impact is above a dollar threshold.

Kill switches

Two kill switches let you stop Autopilot fast:

• Global kill switch - halts every Autopilot action across every app. Find it at the top of the dashboard sidebar when any agent is running, or in Settings → Safety. Activating the global switch puts the workspace into Recovery mode: a banner appears across the dashboard, no new Autopilot actions queue, and you can review in-flight executions before they complete.
• Per-agent kill switch - pause a single agent from Agents → pick agent → Pause. Paused agents stop generating recommendations and stop executing Autopilot actions.

Quiet hours

Configure quiet hours from Settings → Quiet hours. While a quiet window is active no Autopilot actions execute (they queue for when the window closes), and notifications batch into a single digest for the next active hour.

Typical setups: nights and weekends, launch windows, vacation. You can set multiple windows per app or workspace-wide; the dashboard topbar shows a purple Quiet hours chip whenever a window is in effect.

Reversibility & auto-rollback

Every Autopilot action is shipped with rollback metadata. The agent that originally proposed it watches the post-deploy metric (paywall conversion, ASO rank, retention, etc.) and automatically rolls back if the lift falls below a defined threshold within the watch window. The thresholds are visible on the recommendation’s detail page so you know exactly what the safety net is.

Manual rollback is one click from the recommendation detail page or from Agents run history. Rollbacks themselves are logged as their own run.

Notifications

Every Autopilot execution sends a notification immediately (in-app and optionally in Slack). The notification includes the action taken, the rollback window, and a deep link to the run trace. You can subscribe a specific email or Slack channel to Autopilot events only - useful for compliance audits or keeping a co-founder in the loop without flooding them with raw recommendations.

Recovery mode

Recovery mode is a soft reset. While active:

• All Autopilot toggles are forced off, regardless of saved state.
• No new Autopilot actions execute or queue.
• In-flight executions complete normally - they were already approved and rolling back mid-flight is riskier than letting them finish.
• A red banner sits at the top of every dashboard page. The banner links to the kill-switch event log so you can see why recovery mode was activated.

Exit recovery mode from the banner. The previous Autopilot configuration is restored exactly - toggles, thresholds, quiet hours.