Skip to main content

Get started with Automations

Create a saved prompt, run it once, and read the automation run. For what an automation, a trigger, and an automation run are, see Automations.

Open Automations in the sidebar.

Before you begin​

You need:

Only the owner can edit an automation. A project member who can open it can start a run and read the run. For someone else, the Edit button shows Only the owner can edit this automation.

Create an automation​

  1. Open Automations.
  2. Click New automation.
  3. Enter a Name.
  4. Choose a Trigger. Select Manual to run the automation only when someone clicks Run. See Attach a trigger.
  5. Write the Prompt. This prompt applies to every automation run. See Write the prompt.
  6. Choose a Model: Auto, Claude Sonnet 5.5, or Claude Opus 5.5. That model is used for every run. A new automation without a saved model uses the organization's default.
  7. Under MCP servers, choose None, Read-only, or All tools for each server that is turned on and connected. All tools asks you to confirm, then runs writes with no further confirmation. Project members who can start a run can trigger the tools you allowed, including with their own input. To add or connect a server, click Manage external MCPs. A saved server that becomes unavailable stays until you click Remove.
  8. Click Save.

The run calls those servers with your connection, including when another project member clicks Run. See External MCP accounts and services.

Write the prompt​

Say what to look for and what the automation run should report. Name the event type, topic, or threshold when you have one. A short prompt that states the outcome is easier to reuse than a prompt that narrates the investigation.

When you write the prompt:

  • Give the agent context: what key topics mean, the team's failure categories, and a layout that shows the relevant data.
  • Say what to report: timestamps, supporting topics or images, confidence, and any missing data.
  • Separate what was observed from possible explanations, and ask the agent to mark uncertain cases instead of guessing.
  • Split large jobs into smaller chained automations instead of one long prompt. See Choose what triggers a run.
  • Iterate on a handful of recordings where you already know the answer before attaching a trigger.

A disengagement triage prompt:

For the disengagement typed event that was created, open the Disengagement Review layout. Inspect the seconds before and after the event on the layout's topics. Report the topic or fault that best explains why control was handed back, how confident you are, and the time range that supports it.

A prompt that creates a typed event for each latency spike:

The run creates these events when it can open the recording in Context.

Find each interval where /control/loop_latency_ms is above 200 ms in the recording in the context. Open the Control Latency layout to inspect the signal. Create one latency_spike typed event per interval, from the first to the last sample above the threshold. Set properties topic and peak_latency_ms. Do not change existing events. Report each interval and its peak.

A grasp-review prompt:

Review the grasp events in Context using available images and signals. For each event, report the observed outcome, timestamped evidence, other possible explanations, and any missing data. Mark uncertain cases explicitly. Don't infer success from gripper state alone.

Run it once​

  1. Open the automation.
  2. Click Run.
  3. Enter Context. This is the prompt for this automation run, such as recording IDs, an event to inspect, or extra instructions. Context doesn't change the saved prompt. The dialog requires it.
  4. Click Run automation.
  5. On the Automation runs tab, open the new automation run.

The automation run page shows Automation run details, Output, and Activity. Automation run details lists Status, Started, Duration, Token usage, Model, Triggered by, Automation prompt, and Run context. Status is:

  • Pending — the run is waiting to start.
  • Running — the agent is working.
  • Finished — the run completed. It is green when the outcome is clean, and yellow with a warning icon when the agent isn't sure whether something went wrong. The message appears next to the status.
  • Error — a system failure.
  • Failed — the task did not get done, or access is missing.

Output appears after the automation run finishes. It is the agent's final response, or No output available. when the run ended without one. Activity is the steps the agent took, including recordings, topics, and layouts it opened. It is open while the automation run is pending or running, and closed after it finishes.

Attach a trigger​

On the automation, click Edit and choose a Trigger. Select Manual to run the automation only when someone clicks Run.

Trigger can be Manual, On a schedule, Event is created, Event is updated, Recording is created, Recording is imported, or Session is created. For an event trigger, set Event type to one event type or Any event type. Click Save to keep the trigger. See Triggers for what each one means.

For On a schedule, choose Repeat: every hour, day, week, or month. Hourly runs start at the top of the hour. For other cadences, select Hour (UTC), then the day of the week or month when applicable. Monthly dates are limited to days 1–28. Runs delayed by five minutes or more are skipped. Click Save to keep the schedule.

Tips and tricks​

Successful prompts reference layouts, topics, user scripts, and typed events. Name the layout that shows the relevant panels, the topics and event types to inspect, and any user scripts that transform the data.