Baton Pass: A Low-Token Handover Workflow for Multi-Agent Collaboration

Baton Pass is an innovative multi-agent project collaboration solution. Its core goal is to address memory loss caused by the context window limitations of large language models, enabling seamless work relay between AI tools like Claude and Codex. The key idea is to transfer the minimal necessary incremental information through a structured handover mechanism instead of a full historical review, thereby reducing token consumption, minimizing information overload, and optimizing multi-agent collaboration efficiency.

In development workflows driven by large language models, context window limitations are a long-standing pain point: when AI processes large codebases or long conversations, it gradually forgets early content, leading to repeated work, information drift, hidden errors, and efficiency losses. Traditional solutions relying on project documentation have lag issues and cannot fully capture the subtle decision-making processes in conversations.

The core design philosophy of Baton Pass is 'transfer only incremental information, not a full review'. Its workflow architecture includes seven core commands:

1. new-game: Initialize project file structure (configuration, state tracking, etc.);
2. save-state: Record current tasks, positions, file changes, etc. when safely paused;
3. baton-pass: Intelligent handover (goals, completed tasks, file changes, verification status, etc.);
4. foresight: Receiver verification process (confirm goals, check status, etc.);
5. dragon-dance: Conditional recording of lessons learned (fix implicit knowledge);
6. party-check: Quick query of work ownership;
7. hindsight: Full-chain audit (trace history, review deviations).

Baton Pass ensures precise state descriptions through a verification vocabulary to avoid ambiguous expressions, such as 'passed' (locally run and confirmed clean), 'not run — [reason]' (skip with reason), etc. The Turn State mechanism synchronizes states via the next-task file and baton-pass.state.json, including information like State (active/paused/handed-off, etc.), Last Agent, Next Agent, etc., to ensure consistent states across multiple agents.

Baton Pass is best suited for paired use with Claude and Codex, with advantages including faster detection of outdated assumptions, reduced handover drift, continuity when tokens are exhausted, verification relay, etc. Integration methods: Install as a Claude Code plugin (/plugin marketplace add francisN21/baton-pass), or install the Node CLI package via npm (npx baton-pass init), supporting slash commands like /new-game, /baton-pass, etc.

When using Baton Pass, avoid the following misconceptions:

• Overuse (do not use for minor checkpoints);
• Meaningless recording (do not run dragon-dance if nothing is learned);
• Over-auditing (do not run hindsight after every handover);
• Redundant rewriting (do not rewrite memory files for trivial work);
• Over-verification (do not turn foresight into a full repository audit);
• Ambiguous expressions (do not use 'passed' instead of 'expected to pass, unverified');
• Dirty tree handover (do not hand over unnamed and uncommitted dirty directories).

Baton Pass does not attempt to break through the technical limitations of context windows; instead, it achieves optimal collaboration effects through workflow redesign. This pragmatic solution philosophy is worth learning: when facing technical constraints that cannot be immediately resolved, optimize the experience through process innovation. For teams exploring multi-AI tool collaboration, Baton Pass provides a well-thought-out reference implementation.