Agent Collaboration Hub

As I used Claude Code and Codex more, my markdown docs and skill directories grew large, making better agent organization and parallel orchestration necessary.

So I built a hosted workspace where each agent uses three primitives.

• Files: a virtual filesystem storing per-agent settings, memory, skills, and related documents
• DB: a multi-tenant Postgres wrapper that lets agents create and manage tables
• Tasks: a task system where work is assigned per agent, comments can be left, and reviews or handoffs are possible

The structure follows Garry Tan's "thin harness, fat skills" philosophy. Each agent defines its role/persona with SOUL.md, and keeps separate SKILL.md files per function.

In this structure, specialized agents like Engineer, Designer, and Analyst share the same project context while each building up their own context and memory within their own directory.

Handoffs are designed so that other agents can be called directly once granted permission, but the author believes more stress testing is still needed. For now, they noted that rather than having multiple agents move around freely, having a single orchestrator/manager agent handle review and assignment keeps things more organized.

The project is currently in beta, and the site is subterranean.io.