Autopilot: A Detachable Agent Orchestration Framework for AI-Driven Development Workflows

Autopilot is a project-agnostic agent orchestration scaffold designed to address issues such as task conflicts, opaque status, work isolation, and lifecycle management when multiple AI programming assistants collaborate. It coordinates parallel collaboration of multiple AI assistants on codebases through mechanisms like structured TOML to-do lists (Track and Slice), atomic task claiming, and worktree lifecycle management, establishing a predictable and auditable AI-driven development workflow.

With the popularity of AI programming assistants like Claude Code and GitHub Copilot, developers face challenges in multi-AI session collaboration: 1. Conflicts caused by multiple AI sessions modifying the same file; 2. Difficulty in tracking the progress of each AI session; 3. Lack of lightweight branch/worktree isolation mechanisms; 4. No automated environment preparation and cleanup when AI sessions start or stop. Autopilot is designed to address these issues, providing a detachable agent orchestration tool.

Autopilot's core mechanisms include: 1. Track and Slice: TOML defines major functional directions (Track) and specific task units (Slice); 2. Atomic Task Claiming: Prevents conflicts based on file locks, creating branches and worktrees after successful claiming; 3. Worktree Lifecycle: Automatic management from claiming, development to completion/abandonment/expiration; 4. Heartbeat Mechanism: Regularly sends heartbeats to release expired claims; 5. Pre-merge Gating: Runs configured commands (e.g., tests) before PR merging, compatible with branch protection.

Autopilot provides multi-level supervision: 1. Supervision Levels: max (auto-merge), default (decision based on risk), careful (requires maintainer review), cautious (draft PR); 2. Slice Risk Declaration: Each Slice can be marked with low/medium/high risk; 3. Observation Points: Configure conditions to trigger reviews (e.g., new dependencies added, changes to specific paths, file renaming).

State is stored outside the project (~/.local/share/autopilot/...), including state.json (claim/block information) and events.log (audit log). The benefits are no pollution to the project repository and reconstructability. Adapts to various AI assistants: Claude Code can be wrapped as a command, Codex registered via plugins, and Cursor requires copying AGENTS.md and creating symbolic links.

Installation: Clone the repository → pnpm install → link to PATH. Project Initialization: autopilot init to create configuration → autopilot track new to add Track → edit Slice. Daily Workflow: list to view to-dos → claim to take tasks → heartbeat to send heartbeats → prepare-merge to prepare for merging → complete to finish tasks.

Practical Significance: Supports multi-AI session coordination, human-machine collaboration (GitHub Issues + Ledger), progressive adoption, and auditability. Limitations: Depends on Node.js environment, targets squash-merge workflows, has a learning curve, and some features rely on GitHub API.

Autopilot does not replace existing version control or project management tools; instead, it provides a lightweight orchestration layer on top of them to solve multi-AI collaboration issues. As AI programming assistants become more popular, such coordination tools will become increasingly important, laying the foundation for AI-driven large-scale software development.