Multi-Agent Orchestration for Knowledge Workers

How to design a system where multiple AI agents coordinate across your tools — Obsidian, email, calendar, code.

Why One Agent Isn't Enough

A single agent can read your files, answer questions, and take actions. That's powerful. But it has the same limitation as a single employee — it can only do one thing at a time, and it can only know what it directly encounters.

Multiple agents, each with a specialty and a defined scope, create something qualitatively different. They divide cognitive labor the way teams divide task labor. The intelligence analyst doesn't need to also be the content writer. The project manager doesn't need to also be the knowledge graph operator.

The Five-Agent Stack

This is the system I run daily. It's not theoretical — these agents coordinate across my Obsidian vault, terminal, browser, and external tools.

MetaTotem — The Project Manager. Persistent memory across sessions. Tracks projects, people, decisions, and context. Manages the Obsidian vault — file operations, Bases databases, template creation. When I start a new session, MetaTotem knows what I was working on, what's pending, and what's changed since yesterday.

Trailblazer — The Workflow Capturer. Records actions during flow states via Playwright browser automation. Models how I actually work — the sequence of tools, context switches, and information sources I consult. Handles scheduling and coordination with external systems.

Scout — The Analyst. InfraNodus integration for knowledge graph operations. Gap finding, associative trail mapping, competitive intelligence, negative space analysis. When I need to understand the structure of a knowledge domain — not just what's in it, but what's missing — Scout runs the analysis.

3x3 — The Framework Engine. Three perspectives, three scales, three timeframes applied to any analytical question. Produces structured analyses that capture multiple angles simultaneously. Prevents the tunnel vision that comes from analyzing anything from a single vantage point.

ContentFactory — The Publisher. Takes source material — session logs, research findings, analysis outputs — and produces platform-specific content. LinkedIn posts, blog drafts, Twitter threads. Runs anti-slop detection. Formats for each platform's constraints.

The Coordination Layer

The agents coordinate through the simplest possible mechanism: the Obsidian vault file system.

No custom protocols. No message buses. No API orchestration layers. Just markdown files with YAML frontmatter that any agent can read and write.

MetaTotem writes a session log. Trailblazer reads it and captures the workflow pattern. Scout reads the workflow output and runs a gap analysis on the knowledge domain. 3x3 structures the gap analysis into a framework. ContentFactory takes the framework and produces a LinkedIn post.

Each agent reads files, writes files, and follows its CLAUDE.md instructions. The vault is the shared state. Git provides version control. The coordination emerges from the file system, not from an orchestration engine.

Building Your Own Stack

You don't start with five agents. You start with one — your first Claude Code agent with a good CLAUDE.md. Then you add a second for a specific task you find yourself repeating. Then a third.

The pattern: identify a cognitive task you do repeatedly, define it clearly enough that an agent can handle it, create the CLAUDE.md that scopes the agent's role, and connect it to the tools it needs via MCP servers.

Over time, the stack compounds. Each agent makes the others more effective. MetaTotem's memory makes Scout's analysis more targeted. Scout's analysis makes ContentFactory's output more substantive. ContentFactory's output creates artifacts that MetaTotem tracks and Trailblazer captures.

This isn't science fiction. This is Tuesday.