Automations

Each scheduled run creates an Operator task in the project. The automation stores its instructions, schedules, timezone, and repositories to clone. Its run history links to the resulting conversations, so you can inspect the work or continue with Operator afterward.

Before you create an automation

Operator automations currently require an active organization Slack connection. Open Automations in your project and select Connect on the Slack card if Slack is not connected. An organization administrator can also connect Slack from Organization → Integrations.

To receive a Slack notification, name the destination channel in the automation’s instructions and ensure the bot has access to it. An active organization Slack connection is sufficient for sending notifications. You don’t need to bind the destination channel to the project. See Operator in Slack for connection setup and the separate channel binding used for inbound Operator conversations.

If the work needs source code, register the repository through the project’s GitHub integration. You can select registered repositories to clone when configuring the automation.

Create a daily briefing

1. Open Automations and select the Daily briefing template. The template fills in a name, instructions, and a daily schedule.
2. Read Operator instructions and adjust the scope. Include the full Slack channel name if you want the result sent there. Replace any runtime or reviewer placeholders when using a different template.
3. Under Triggers, choose the recurrence and time. Check the displayed timezone, which comes from your browser when you create the automation.
4. Review Cloned repositories. Keep the repositories needed for this work, or leave the selection empty for a report that only reads project evidence.
5. Leave the switch Active to enable the schedule, then select Create.

For a custom schedule and prompt, select New automation instead of a template, complete the same fields, and select Save. Templates only prefill the form. After creation, you can edit their instructions and triggers like any other automation.

Choose a template

Each template fills in a name, Operator instructions, and a schedule. Choose the template that matches the work you want Operator to repeat, then review its instructions and schedule. Add any runtime slug, team member email, or Slack destination requested by the template. Select a registered repository when the work reads or changes Recipe code.

Project check-in

Use this template when Operator already has open work in the project. Operator checks its issues, pull requests, and expert review assignments. It continues work after a review or human answer and reports what still needs attention. If the work uses a repository, select it when you create the automation.

Daily briefing

Use this template for a daily summary of activity across the project’s runtimes. Operator reports changes in traffic, judge pass rates, feedback, patterns, and work waiting on a person. For a report based only on project evidence, leave Cloned repositories empty.

Build expert review queue

Use this template to give experts a small, varied set of recent conversations to review. Operator selects representative failures, disagreements, and normal conversations for comparison, then labels the relevant spans needs-review. The project needs recent conversation activity for Operator to build a useful queue.

Triage judge failures

Use this template when recurring judge failures need input from a domain expert. Operator groups related failures, selects representative spans, and assigns them to the expert with a specific question. Replace (expert email) and narrow the set of judges if needed.

Grow the eval set

Use this template to turn reviewed production cases into evaluation coverage. Operator adds cases from annotations whose human comments explain the expected behavior, tests the Recipe update, and prepares at most one draft pull request. Replace (runtime slug) and (engineer email), then select the registered Recipe repository.

Recalibrate judges

Use this template to check whether judge results still agree with human reviews. Operator builds calibration cases from reviewed conversations and compares the results with the previous calibration. It can prepare a draft rubric change or send unresolved disagreements to a domain expert. Replace (runtime slug) and (expert email). Select the registered Recipe repository if Operator may prepare a pull request.

Cost and loop watch

Use this template to find unusual changes in cost, tool use, or latency. Operator compares recent activity with the preceding baseline, inspects representative outliers, and reports whether they suggest a retry loop, a failing tool, or increased traffic. The project needs enough recent activity to establish a useful baseline.

Recipe PR review

Use this template to check new Recipe pull requests before your team reviews or merges them. Operator looks for wider tool permissions, new external calls, and instructions that trust fetched content. It leaves comments on the relevant lines while your team retains responsibility for approval and code changes. Select the registered Recipe repositories you want Operator to review.

Write instructions for an unattended run

Give a scheduled run enough detail to work without asking you questions. Specify the runtime or project area, the time period to investigate, the action to take, and what should happen when there is nothing useful to report. Include a full Slack channel name when the action includes a notification.

Operator checks existing work before starting another change. Its scheduled workflow limits each run to at most one pull request. When the next step needs a person, it records what is blocked and who needs to respond rather than waiting for an answer inside the scheduled run. State the relevant person’s email in advance for workflows that assign review or follow-up work.

Choose a narrow prompt when you want a narrow action. For example, “report cost outliers” asks for a report; “prepare a tested recipe change and open a draft pull request” also requests repository work. Operator’s project access is the same access described in Operator.

Schedules and conditions

The trigger editor supports daily, weekly, monthly, and hourly schedules. Select Add Trigger to add another schedule to the same automation, such as Monday morning and Friday afternoon. Introspection uses the earliest next occurrence across the schedules and evaluates them in the saved timezone. Check Next run after saving changes.

Automations created in the dashboard include a condition that checks whether an earlier task from the same automation is still running. If so, the new attempt is cancelled before Operator starts. The same condition applies to manual runs. Separate automations can run at the same time.

The API can store additional conditions, such as requiring new tasks since the previous run or resolved issues from the latest runtime review. Conditions belong to the saved automation. Instructions such as “report only when something changed” are evaluated by Operator while it does the work. See the Data Plane API schema for automation endpoints and fields.

Share an automation

An automation is private to its creator and administrators until you share it. Open Sharing and select member tags to let members with a matching tag read, edit, and run the automation. They can also open its past runs and generated output files.

Removing a tag removes the access granted by that tag. Separate shares remain, and deleting an automation retains its sharing until you revoke it.

If you send a report to Slack, make sure its recipients can also open the linked runs and output files. Use the dashboard’s sharing controls to grant access.

Run, pause, and inspect

Open an automation to edit its Settings or view Run history. Save instruction or trigger changes before starting a manual run.

Select Run to execute the active automation now. A manual run uses the saved instructions and conditions, and the next scheduled occurrence is recomputed afterward. The dashboard disables Run while the automation is inactive. Both manual and scheduled Operator runs require the organization Slack connection to remain active.

Turn the switch from Active to Inactive to pause future scheduled runs. Pausing does not stop work already in progress. While runs are active, select Stop all runs to cancel their current work.

Run history shows each recorded task’s trigger, repositories, start time, status, and duration. Select a row to open its Operator conversation. A cancelled attempt can mean that its conditions were not met, so inspect the reason before treating it as an execution failure.

When Operator finds nothing to do, its scheduled workflow records a quiet run and answers NO_REPLY. Quiet runs are hidden in history by default. Select the Show control for quiet runs to inspect them. You can read the result in the conversation whether or not you requested a Slack message.

Related

• Operator explains how the AI engineering agent investigates production evidence, tests Recipe changes, and prepares pull requests.
• Operator in Slack explains Slack connections, channel bindings, and message destinations.
• Learn from production explains the evidence behind review and improvement work.