Automations

What Are Automations?

Automations are unattended agent runs — AI agents that execute on a schedule or in response to events, without a human in the loop. Each automation configures:

• Agent — which Virtual MCP (agent) to run
• Messages — the prompt/instructions the agent receives each run
• Model — which AI model and credentials to use
• Triggers — when the automation fires (cron schedule, connection event, or manual)

Automations run in passthrough mode, meaning all tool calls are auto-approved. This is what makes them unattended — but it also means you should scope agent tools carefully.

Automations vs Decopilot: Decopilot is interactive — a human converses with the agent and approves actions. Automations are fire-and-forget: the agent runs, completes its task, and the result appears in Monitoring.

When to Use Automations

• Scheduled tasks — daily reports, periodic data syncs, cleanup jobs
• Event-driven workflows — new order → fulfillment, new ticket → triage, deployment → notification
• Any workflow that should run without human approval

Creating an Automation

1. Navigate to Automations in the sidebar
2. Click New Automation
3. Choose an agent — the Virtual MCP that will execute
4. Write the prompt/messages — what the agent should do each run
5. Select a model and credentials
6. Set temperature (default 0.5)
7. Add one or more triggers (cron or event)
8. Activate the automation

Triggers

Cron Triggers

Cron triggers run the automation on a recurring schedule using standard cron expression syntax. All times are in UTC.

Minimum interval: 60 seconds. Expressions that resolve to intervals shorter than 60 seconds are rejected.

Crash-safe scheduling: The system records last_run_at for each trigger. If the server restarts, missed runs are detected and fired on the next poll cycle (every 30 seconds).

Event Triggers

Event triggers fire the automation when a connected MCP server publishes a matching event.

• The connection must implement TRIGGER_CONFIGURE (so the mesh can register/unregister interest in events)
• You specify the connection, event type, and optional params for filtering
• Fail-atomic: if configuring the trigger on the connection fails, the trigger is not created
• When the automation is deactivated or deleted, event triggers are disabled on their connections (best-effort)

Param filtering: If you specify params, only events whose data contains all matching key-value pairs will fire the automation. Extra fields in the event data are ignored.

Running and Testing

• Use the AUTOMATION_RUN tool or the UI to manually trigger a run at any time
• Each run creates a thread visible in Monitoring
• Runs that error are marked as failed in the thread

Safety and Limits

• Per-automation concurrency: Maximum 3 concurrent runs per automation. Additional triggers are skipped until a slot opens.
• Global concurrency: Maximum 10 concurrent automation runs across the entire system.
• Run timeout: Each run is capped at 5 minutes.
• Recursion prevention: Event trigger chains are limited to a depth of 3. If an automation’s actions trigger events that fire other automations, the chain stops after 3 levels.
• Payload limit: Event data is capped at 1 MB. Larger payloads are truncated to mitigate prompt injection risks.
• Creator validation: If the automation’s creator leaves the organization, the automation is automatically deactivated.

Automation MCP Tools

These tools are available for managing automations programmatically via the MCP interface:

Next steps: Learn about Agents to understand how to design the Virtual MCPs that automations execute, or explore Monitoring to track automation runs.