cowork-overseer

A template for building a persistent multi-agent AI ecosystem using Claude's Cowork mode and an Obsidian-compatible markdown vault as shared external memory.

One coordinator agent (Overseer) maintains awareness across all projects. Each project has its own specialized sub-agent (Coworker) with its own identity, vault, and operating rules. The coordinator can see everything; each sub-agent sees only its own project.

This is the multi-agent version of cowork-agent. If you have one project or domain, use that instead. This template is for managing several parallel workstreams with cross-project awareness.

The Problem This Solves

The single-agent problem (context compaction, losing memory across sessions) is solved by the vault pattern described in cowork-agent. Read that README first if this concept is new to you.

The multi-agent problem is different: no single AI instance can effectively maintain identity and context across multiple distinct projects simultaneously. If you ask one agent to manage software development, genealogy research, a hardware build, and a blog all in one session, it loses coherence. Context from one domain bleeds into another. The agent can't develop a genuine working relationship with any single domain.

The Overseer pattern solves this with separation: