A new open-source multi-agent orchestrator called lean-collab aims to solve the limitations of single-agent LLMs on complex, long-running tasks. The tool coordinates multiple Claude Code agents to collaborate on theorem proving using Lean 4, addressing common issues like stalling, looping, and generating non-compiling code.

Built around an orchestrator agent that manages task decomposition, the system spawns parallel sub-agents that share real-time context and discoveries. It was tested on Putnam-level math problems but generalizes to refactors, app builds, and research tasks. The architecture uses Ensue Memory Network for coordination and requires Lean 4, Rust, and an Ensue API key.

Users must configure local settings including theorem IDs, parallel agent limits, and Lean project paths. The CLI provides session management, goal tracking, tactic verification, and search capabilities. A warm server significantly speeds up verification by keeping Mathlib loaded in memory. Token usage can be substantial due to parallel agent spawning, so the authors recommend starting with fewer workers and monitoring progress closely.