February 12, 2026
Twenty bots, zero chill
Show HN: 20+ Claude Code agents coordinating on real work (open source)
20+ AI sidekicks team up; hype meets hard questions
TLDR: An open-source tool coordinates 20+ Claude Code agents to tackle formal math proofs together. The crowd is split between fans praising transparency and coordination, skeptics probing failure modes and licensing, and curious power users asking how progress signals work—important because multi-agent workflows could make AI actually reliable.
Meet the latest Show HN flex: an open-source setup that wrangles 20+ Claude Code “agents” to work together on real math proofs, using the Lean 4 toolchain and an Ensue memory network. There’s a Rust command-line app, a “warm server” to speed checks, and lots of knobs to tune. Translation: it's an AI team project manager, not just one bot.
Commenters instantly split into camps. clairekart wants the receipts, asking if single-bot Claude fails by “looping, context drift, plan collapse, tool misuse.” miligauss cheers the transparency: Claude alone is a black box, but here you can see the plan and mistakes; no more “Ralph” doing the infinite loop dance. Meanwhile yodon slams the brakes: open source without a clear license? MIT vs GPL v3 isn’t vibes, it’s law.
Others are in try-it-now mode. christinetyip asks for a starter challenge where this swarm beats one bot. zmanian digs into the live updates: do agents watch test failures and build status, or just a global feed? The memes are strong—“AI sweatshop” vs “AI orchestra”—but the mood is cautiously excited: if multi-agent beats chaos and shows its work, people will actually use it. Repo is here: github.com/anthropics/lean-collab-plugin, API info at ensue.dev.
Key Points
- •The system enables multi-agent collaborative theorem proving in Lean 4 via the Ensue Memory Network.
- •A Rust-based CLI (lc) coordinates sessions, goals, and tactic verification using Lean’s LSP.
- •Setup requires Lean 4 with Mathlib, Rust toolchain, and an Ensue API key; configuration is done in .lean-collab.json.
- •A warm server preloads Mathlib, reducing tactic verification time from ~20s to ~2–5s.
- •CLI supports goal lifecycle management (claim, verify, decompose, backtrack, compose) and search/exploration commands.