MassGen

🚀 MassGen: Multi-Agent Scaling System for GenAI

MassGen is a cutting-edge multi-agent system that leverages the power of collaborative AI to solve complex tasks.

Scaling AI with collaborative, continuously improving agents (4x speed)

MassGen is a cutting-edge multi-agent framework that coordinates AI agents to solve complex tasks through redundancy and iterative refinement. Every agent tackles the full problem, observing, critiquing, and building on each other's work across cycles of refinement and restarts. When agents believe there is a strong enough answer, they vote, and the best collectively validated answer wins. This approach to parallel refinement and collective validation lays the groundwork for principled multi-agent scaling, where the system continuously improves its outputs by leveraging diverse agent perspectives and enforcing quality through consensus.

This project started with the "threads of thought" and "iterative refinement" ideas presented in The Myth of Reasoning, and extends the classic "multi-agent conversation" idea in AG2. Here is a video recording of the background context introduction presented at the Berkeley Agentic AI Summit 2025.

🧩 Use MassGen as a Skill: npx skills add massgen/skills --all — then type invoke the skill in Claude Code, Cursor, Copilot, or 40+ other agents. Learn more →

📚 For Contributors: See MassGen Contributor Handbook - Centralized policies and resources for development and research teams

---

📋 Table of Contents

✨ Key Features

• Cross-Model/Agent Synergy
• Parallel Processing
• Intelligence Sharing
• Consensus Building
• Live Visualization

🆕 Latest Features

• v0.1.96 Features

🏗️ System Design

• System Architecture
• Parallel Processing
• Real-time Collaboration
• Convergence Detection
• Adaptive Coordination

🚀 Quick Start

• 📥 Installation
• 🔐 API Configuration
• 🧩 Supported Models and Tools
  • Models
  • Tools
• 🏃 Run MassGen
  • CLI Configuration Parameters
  • 1. Single Agent (Easiest Start)
  • 2. Multi-Agent Collaboration (Recommended)
  • 3. Model Context Protocol (MCP)
  • 4. File System Operations
  • 5. Project Integration (NEW in v0.0.21)
  • Backend Configuration Reference
  • Interactive Multi-Turn Mode
• 📊 View Results
  • Real-time Display
  • Comprehensive Logging

🤖 Automation & LLM Integration

• Automation Mode
• BackgroundShellManager
• Status File Reference
• Full Automation Guide

💡 Case Studies & Examples

• Case Studies

🗺️ Roadmap

• Recent Achievements (v0.1.97)
• Previous Achievements (v0.0.3 - v0.1.96)
• Key Future Enhancements
  • Bug Fixes & Backend Improvements
  • Advanced Agent Collaboration
  • Expanded Model, Tool & Agent Integrations
  • Improved Performance & Scalability
  • Enhanced Developer Experience
• v0.1.97 Roadmap

📚 Additional Resources

• 🤝 Contributing
• 📄 License
• ⭐ Star History

---

✨ Key Features

Feature	Description
🤝 Cross-Model/Agent Synergy	Harness strengths from diverse frontier model-powered agents
⚡ Parallel Processing	Multiple agents tackle problems simultaneously
👥 Intelligence Sharing	Agents share and learn from each other's work
🔄 Consensus Building	Natural convergence through collaborative refinement
🖥️ Live Visualization	Interactive Textual TUI with timeline, agent cards, and vote tracking (default). Also available: Web UI, Rich display.

---

🆕 Latest Features (v0.1.97)

🎉 Released: June 12, 2026

What's New in v0.1.97 (Application-Layer Permission Engine):

• 🛡️ Layered Permission Engine - Opt-in permissions: block routes every tool call through a non-overridable hardline floor (rm -rf /, fork bombs), declarative allow/ask/deny rules over a small action(target) algebra (deny-wins), and a blast-radius risk classifier — auto-allowing reads/in-workspace edits and asking only for the dangerous tail (egress, force-push, publish, privilege). The app-layer companion to v0.1.96's OS sandbox.
• ✋ Approval That Fits the Run - An ask resolves via an automation policy (risk-based / deny-all / allow-all) or a file request/response handshake for headless/remote approval (Slack bot, /approve <id>, …) — fail-closed by design.
• 🧑‍🤝‍🧑 Roles, Audit & Guards - Per-agent role presets (e.g. read-only, which also empties the agent's OS-sandbox writable set), an append-only JSONL audit ledger of every decision, a runaway-loop budget, always-grant persistence, and a channel-based guardrail prompt that nudges the model to surface blocks rather than circumvent them while keeping ask sanctioned