A multi-agent system framework designed to tackle highly complex, long-duration problems requiring sustained computational effort, such as compiler design or mathematical theorem proving. It targets users needing to solve intricate, abstract, and verifiable tasks, offering a structured approach to continuous AI-driven problem-solving over extended periods, potentially exceeding 100 hours.

How It Works

This framework employs a Master Agent and Subagent architecture. The Master Agent defines an overarching Goal and specific Criteria for success. It then instantiates a Subagent tasked with persistently working towards achieving the Goal. The Master Agent periodically monitors the Subagent's activity and evaluates its progress against the defined Criteria. If the Criteria are unmet or the Subagent becomes inactive, the Master Agent automatically restarts a new Subagent to continue the task, ensuring uninterrupted progress until the Goal is successfully met.

Quick Start & Requirements

Usage involves copying a provided prompt template into a multi-agent supported tool (e.g., Claude Code, Codex, OpenClaw). Users must meticulously define their specific Goal and Criteria for success within this template. The system may consume significant LLM tokens and computational time. Links to example projects, such as a TypeScript compiler in C++ and SQLite in Rust, are provided for reference.

Highlighted Details

• Successfully developed a TypeScript compiler in C++ within approximately 100 hours.
• Developed SQLite in Rust in approximately 30 hours.
• Designed for sustained operation exceeding 300 hours on complex problems.

Maintenance & Community

No specific details regarding maintenance, notable contributors, community channels (e.g., Discord, Slack), or roadmaps are present in the provided text.

Licensing & Compatibility

No license information is provided in the README content.

Limitations & Caveats

This system requires an external multi-agent execution environment. Processes can consume substantial LLM tokens and time, necessitating adequate API plan or subscription resources. Users are cautioned against integrating the prompt template directly into AI agent skills or plugins to prevent potential context window contamination. The framework relies on third-party tools for agent execution.