The Baton Pattern

A lightweight handoff protocol for multi-agent AI pipelines. A small JSON object that carries context between workflow stages.

How the baton flows

1. 1.Planner → baton + patch
2. 2.Implementer → baton + patch
3. 3.Tester → baton + patch
4. 4.Reviewer → baton + patch

Example Structure (~1,000 tokens)

{
  "goal": "Add user auth",
  "current_state": ["Routes created", "Tests passing"],
  "decision_log": ["Chose JWT over sessions"],
  "constraints": ["No breaking changes"],
  "work_scope": ["src/auth/*"],
  "artifacts": [],
  "open_questions": [],
  "acceptance": ["All auth tests green"]
}

The Problem

When multiple AI agents collaborate sequentially—planning, building, and reviewing—they struggle with context transfer. Key challenges include:

• >Passing full artifacts (code, docs) between every stage costs too many tokens
• >Even with complete artifacts, agents overlook the reasoning behind decisions

Without structured handoffs, agents duplicate effort or contradict prior choices.

The Solution

The baton is a small JSON object (~1,000 tokens) that travels between stages. Each agent reads it before starting and patches it upon completion—similar to a relay race baton carrying notes.

Structure

How It Works

1. 1.Initialize: The baton begins with a goal and initial state
2. 2.Inject: Before each stage, inject the baton into the agent's prompt as markdown
3. 3.Patch: After executing, the agent returns a baton_patch—only changed fields update
4. 4.Repeat: The updated baton feeds into the next stage until completion

Baton vs. Artifacts

The baton communicates "what happened and why," while artifacts represent the actual work product.

Key Design Choices

• >Append-only decisions: Earlier decisions persist, preventing contradictions
• >Replace semantics for state: current_state reflects current truth, not history
• >Budget-friendly: At ~1,000 tokens, the baton always fits in context
• >Full audit trail: Every patch records timestamp and stage ID

When to Use This

The baton pattern suits any multi-step AI pipeline where:

• >Agents need context from previous steps
• >Token budgets are limited
• >Decisions must remain consistent across stages
• >Context evolution requires tracing

It maintains intentional simplicity—just a JSON object with merge-patch updates.