huashu-agent-swarm

Infinite Agent Loop - Infinite Agent Swarm Mode

Inspired by Nicholas Carlini autonomously building a C compiler with 16 Claude instances. There is no master agent, pure git self-organization, each agent independently claims tasks, writes code, and pushes.

Trigger Conditions

Use this skill when the user mentions "swarm mode", "multi-agent parallel", "infinite loop", "agent swarm", "start swarm".

Prerequisites

• tmux（brew install tmux）
• claude CLI (already installed)
• git repository (existing or newly created)

Usage Flow

Step 1: Describe the project

The user tells me:

• Project directory path (must be a git repository)
• Project goals and overall description
• Initial task list (or let the agent break it down itself)
• Number of agents (default 8)
• Code standards and test commands

Step 2: Initialize the project

bash SKILL_DIR/scripts/setup_project.sh <项目目录>

This will create in the project:

• AGENT_PROMPT.md - Generated from a template, needs me to customize it according to user requirements
• TASKS.md - Initial task list
• current_tasks/ - Task claiming directory
• agent_logs/ - Log directory

Then I customize references/agent-prompt-template.md based on AGENT_PROMPT.md, filling in the project-specific information.

Step 3: Start the swarm

bash SKILL_DIR/scripts/start_swarm.sh <agent数量> <项目目录>

This will:

1. Create a git worktree for each agent (share the .git object store, no disk waste)
2. Create a tmux session, one agent per pane
3. Each agent enters an infinite loop: pull → claim task → execute → push → next

Step 4: Open the observation dashboard

python3 SKILL_DIR/scripts/dashboard.py <项目目录> 8420

Open in browser http://localhost:8420，可以：

• View the status of all agents, git log, and task progress in real time
• View the latest log of each agent
• Send commands directly to agents via the input box (written to HUMAN_INPUT.md)
• Stop all agents with one click

You can also monitor via the command line:

# 终端状态
bash SKILL_DIR/scripts/status.sh <项目目录>

# 发送指令
bash SKILL_DIR/scripts/send_input.sh <项目目录> "你的指令"

# 直接进入tmux观察
tmux attach -t swarm-<项目名>

Step 5: Stop

bash SKILL_DIR/scripts/stop_swarm.sh <项目目录>

Automatically stop all agents + merge branches + clean up worktrees.

Core Mechanisms

Git self-organizing coordination

• Each agent claims tasks through current_tasks/*.lock files
• Through TASKS.md to understand global progress
• Through git log to understand the work of other agents
• Conflicts are resolved by the agents themselves

Git Worktree isolation

• No multiple clones, use git worktree to achieve isolation
• All worktrees share the same .git object store
• Each agent works independently in its own worktree

Infinite loop

• Each agent automatically starts the next session after completing one
• Through git pull to get the latest results of other agents
• Through sleep intervals to avoid API rate limiting

Key Configuration

Parameter	Default Value	Description
Number of agents	8	Can be specified at startup
sleep interval	5 seconds	Adjustable in agent_loop.sh
Model	claude-opus-4-6	Adjustable in agent_loop.sh

Risks and Mitigation

Risk	Mitigation
API rate limiting	sleep interval + adjustable number of agents
Merge conflicts	AGENT_PROMPT guides small-grained commits
Infinite loops and useless work	Log monitoring + stop conditions
Disk space	stop_swarm.sh automatic cleanup
Runaway costs	The number of sessions can be limited in AGENT_PROMPT

---

Produced by Hua Shu | AI Native Coder · Indie Developer WeChat Official Account "Hua Shu" | 300K+ followers | AI tools and productivity enhancement Representative work: Xiaomao Fill Light (AppStore Paid Rankings Top 1) · "One Book to Master DeepSeek"