> ## Documentation Index
> Fetch the complete documentation index at: https://docs.trydraft.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Autopilot Mode

> Let the planner autonomously manage your ticket pipeline

## Overview

Draft's **Planner** (autopilot) runs on periodic ticks and autonomously manages the ticket pipeline:

* Picks the highest-priority `PLANNED` ticket and executes it
* Proposes follow-up tickets when tickets become `BLOCKED`
* Generates reflections for completed tickets
* Manages ticket dependencies and unblocking

## Enabling Autopilot

Autopilot is controlled in `draft.yaml`:

```yaml theme={null}
planner_config:
  auto_execute: true
  model: "claude-sonnet-4-20250514"
```

## What the Planner Does

### Every 2 Seconds

The planner tick runs and checks:

1. **Are there `PLANNED` tickets?** → Pick the highest priority one and execute
2. **Are there `BLOCKED` tickets?** → Analyze the failure and propose follow-ups
3. **Are there `DONE` tickets?** → Generate reflections for learning
4. **Are dependencies resolved?** → Auto-unblock tickets whose blockers are done

### Safety Boundaries

The planner is intentionally limited:

| Can Do                      | Cannot Do                |
| --------------------------- | ------------------------ |
| Create `PROPOSED` tickets   | Transition ticket states |
| Enqueue execute/verify jobs | Delete tickets or goals  |
| Add comments to tickets     | Modify existing tickets  |
| Propose follow-ups          | Override human decisions |

### Rate Limits

| Setting                    | Default | Description                              |
| -------------------------- | ------- | ---------------------------------------- |
| `max_followups_per_ticket` | 2       | Max follow-up tickets per blocked ticket |
| `max_followups_per_tick`   | 3       | Max follow-ups created per planner tick  |

## Concurrency Control

The planner uses a `planner_locks` table to ensure only one planner instance runs at a time. This prevents race conditions when multiple processes might try to plan simultaneously.

## Monitoring

Watch the planner's decisions in real-time:

* **Ticket Events** — Every planner action is logged as a ticket event
* **Job Logs** — Execution and verification logs stream in the UI
* **API** — Query planner status via `/planner/status`

## Best Practices

<Tip>
  Start with `auto_execute: false` to manually approve each execution. Once you trust the pipeline, enable autopilot for fully autonomous operation.
</Tip>

* Set cost budgets on goals to prevent runaway spending
* Review `PROPOSED` tickets before they execute
* Monitor the first few automated runs closely
* Use verification commands to catch issues early