Introducing Lean Collab: A Multi-Agent Orchestrator for Long-Running LLM Tasks
Lean Collab is an open-source orchestrator developed by Mutable State Inc., designed to handle complex, long-running tasks typically unsuitable for single-agent large language models (LLMs). The orchestrator breaks tasks down into manageable components, delegating them to sub-agents that work in parallel and share discoveries in real-time.
Key Features
- Task Decomposition: The orchestrator agent breaks down lengthy and intricate tasks into smaller assignments for sub-agents to handle.
- Parallel Sub-Agents: Sub-agents execute their tasks concurrently, speeding up processing times.
- Task State and Progress Subscription: Track the task's progress with real-time updates, allowing for dynamic adjustments as needed.
- Real-Time Intermediate Sharing: Intermediate discoveries among agents are shared in real-time to improve overall task efficiency and accuracy.
This setup has been tested on complex math problems at the Putnam level but also applies to software refactoring, app building, and comprehensive research tasks.
Installation and Setup
Prerequisites include Lean 4 and Mathlib installation. Configure your environment by installing tools like Rust and setting up the Lean project with appropriate dependencies. For authentication, you'll need an API key from ensue.dev. Sample configuration files and environment setup details are provided in the source documentation.
For a detailed setup walkthrough and source code, it's recommended to clone the repository and review the README for instructions on configuring and deploying your own instance of Lean Collab.
📖 Read the full source: HN AI Agents
👀 See Also
Running NemoClaw with Local vLLM: Setup Notes and Agent Engineering Observations
A developer documented running NVIDIA's NemoClaw sandboxed AI agent platform with a local Nemotron 9B v2 model via vLLM on WSL2. Key findings include inference routing details, parser compatibility issues, and observations about the agent engineering gap.
Crime Team: Multi-Agent Orchestrator for OpenClaw — Parallel Code Review with Coder Agent
Crime Team v0.1 runs multiple specialist OpenClaw agents in parallel for code review, then integrates findings. Includes per-agent models, a coder agent that applies changes, and a re-audit loop. CLI + GUI.
Developer builds AI framework with 17 biological principles using Claude Code
A developer created an AI framework called Cognitive Sparks by implementing 17 biological principles like threshold firing and Hebbian plasticity, based on the 1999 book 'Sparks of Genius.' The entire project—22 design docs and 3,300 lines of code—was built in one day using Claude Code, with no human-written code.
Moving from CLAUDE.md rules to infrastructure enforcement with Citadel
A developer found that adding more rules to CLAUDE.md beyond about 100 lines reduced compliance, with 40% redundancy in their file. The solution was moving enforcement from instructions to infrastructure using lifecycle hooks, skills, and campaign files, culminating in the open-source Citadel system.