Show HN: 20+ Claude Code agents coordinating on real work (open source)

Source: Hacker News

Overview

Single‑agent LLMs struggle with long‑running, complex tasks.
We’ve open‑sourced a multi‑agent orchestrator that handles such tasks by allowing agents to coordinate over shared context while work is in progress. Single LLM agents often stall, loop, or produce non‑compiling code, so this harness enables collaborative execution.

How it works

1. Orchestrator agent – manages task decomposition.
2. Sub‑agents – perform parallel work.
3. Subscriptions – monitor task state and progress.
4. Real‑time sharing – intermediate discoveries are shared between agents.

We tested the system on a Putnam‑level math problem, and the pattern generalizes to refactors, app builds, and long‑form research.

The orchestrator is packaged as a Claude Code skill, designed to be small, readable, and modifiable.

Use it, break it, and let us know what workloads we should try next!

Comments on Hacker News (10 points, 18 comments)