anthropics / anthropics/claude-cookbooks
GNAP: add multi-agent coordination recipe using git as task board
- Dominant language
- Jupyter Notebook
- Stars
- 52.7k
- Forks
- 6.3k
- Avg merge
- 25m
- Merged PRs (30d)
- 6
Description
## Recipe request: GNAP pattern for multi-Claude-agent coordination
The Claude Cookbook covers a wide range of capabilities — tool use, RAG, multimodal, customer service agents. One gap: **multi-agent coordination patterns** where multiple Claude instances collaborate on shared tasks without a central orchestrator.
[GNAP](https://github.com/farol-team/gnap) (Git-Native Agent Protocol) would make an excellent cookbook recipe. It's a simple, Claude-native pattern: use a git repo as a shared task board. Tasks move through `board/todo/` → `board/doing/` → `board/done/` via file operations that Claude already knows how to do.
**Proposed recipe: multi-claude-gnap-coordination.ipynb**
```python
import anthropic
import subprocess
client = anthropic.Anthropic()
def claude_agent_worker(agent_name: str, repo_path: str):
"""A Claude agent that claims and completes GNAP tasks."""
# Claim a task from board/todo/
tasks = list_tasks(repo_path, "todo")
if not tasks:
return
task = tasks[0]
claim_task(repo_path, task, agent_name) # move todo/ → doing/
# Process with Claude
task_content = read_task(repo_path, "doing", task)
response = client.messages.create(
model="claude-opus-4-5",
max_tokens=1024,
messages=[{"role": "user", "content": f"Complete this task: {task_content}"}]
)
# Commit result to board/done/
complete_task(repo_path, task, response.content[0].text)
```
This recipe would demonstrate:
1. Multi-agent parallelism without a message broker
2. Crash-safe task claiming via git atomic operations
3. Full audit trail via git history
4. Works with any number of Claude instances
Given the cookbook's focus on practical patterns developers can copy, GNAP is a strong fit — zero dependencies beyond git and the Claude SDK.
Spec: https://github.com/farol-team/gnap
Contributor guide
Assessment
This issue has not been assessed yet.