Show HN: Autonomous Research Swarm – Repo as shared memory for multi-agent AI

1 point

4 months ago

This project was inspired by Cursor's recent work on scaling long-running autonomous coding (https://cursor.com/blog/scaling-agents), where they ran hundreds of concurrent agents to build a web browser from scratch. Their key insight—using a Planner/Worker/Judge pipeline instead of flat agent coordination—resonated with challenges I was facing in my own research workflows.

I built this repo template to solve a specific problem: how do you run multiple AI coding agents (Claude Code, Codex CLI, etc.) on the same research project without coordination chaos?

The core idea: Don't build complex "agents talking to each other" systems. Instead, use the repository itself as shared memory.

How it works:

• Planner creates scoped tasks with explicit ownership boundaries (allowed/disallowed file paths)

• Workers execute tasks in isolated git worktrees—no merge conflicts during execution

• Judge runs deterministic quality gates before marking work complete

• Contract files lock critical definitions (metrics, schemas) to prevent definition drift

The template includes an example empirical research project (Ethereum L2 rollup economics) with task templates, workstream definitions, and an optional supervisor script for unattended overnight runs.

Everything is file-based and version-controlled—no hidden state, no message queues, just Markdown task files that humans can read and review.

Happy to discuss the technical decisions (why git worktrees over branches, why file-based coordination over a proper queue, etc.) and hear how others are approaching multi-agent coordination.