Infinite Agent Loop - Infinite Agent Swarm Mode

Inspired by Nicholas Carlini using 16 Claude instances to independently build a C compiler. No master agent, pure Git self-organization, each agent independently claims tasks, writes code, and pushes.

Trigger Conditions

Prerequisites

• tmux (

  brew install tmux

  )
• claude CLI (installed)
• Git repository (existing or newly created)

Usage Process

Step 1: Describe the Project

• Project directory path (must be a Git repository)
• Project goals and overall description
• Initial task list (or let agents break down tasks automatically)
• Number of agents (default: 8)
• Code specifications and test commands

Step 2: Initialize the Project

bash SKILL_DIR/scripts/setup_project.sh <project directory>

• AGENT_PROMPT.md

  - Generated from template, needs to be customized according to user requirements

• TASKS.md

  - Initial task list

• current_tasks/

  - Task claim directory

• agent_logs/

  - Log directory

AGENT_PROMPT.md

references/agent-prompt-template.md

Step 3: Start the Swarm

bash SKILL_DIR/scripts/start_swarm.sh <number of agents> <project directory>

1. Create a git worktree for each agent (shares .git object library, 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 Dashboard

python3 SKILL_DIR/scripts/dashboard.py <project directory> 8420

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

# Terminal status
bash SKILL_DIR/scripts/status.sh <project directory>

# Send instructions
bash SKILL_DIR/scripts/send_input.sh <project directory> "your instruction"

# Directly enter tmux to observe
tmux attach -t swarm-<project name>

Step 5: Stop

bash SKILL_DIR/scripts/stop_swarm.sh <project directory>

Core Mechanisms

Git Self-Organization Coordination

• Each agent claims tasks via

  current_tasks/*.lock

  files
• Understands global progress through

  TASKS.md

• Learns about other agents' work through

  git log

• Conflicts are resolved by agents themselves

Git Worktree Isolation

• No multiple clones, uses

  git worktree

  for isolation
• All worktrees share the same

  .git

  object library
• Each agent works independently in its own worktree

Infinite Loop

• Each agent automatically starts the next session after completing one
• Gets the latest results from other agents via

  git pull

• Uses sleep intervals to avoid API rate limits

Key Configurations

Risks and Countermeasures

Produced by Uncle Hua | AI Native Coder · Independent Developer Official Account "Uncle Hua" | 300,000+ followers | AI Tools & Efficiency Improvement Masterpieces: Kitten Fill Light (Top 1 in AppStore Paid Charts) · "Mastering DeepSeek in One Book"