Multi-Agent Architecture for Personal AI

This article is part of a series about MACX Exocortex.

The Problem With Single-Context AI

[TODO: Context pollution, the 200K token limit, why you can’t just have one big conversation]

Hub and Spoke

[TODO: Orchestrator coordinates, workers execute, results flow back]

       Fred
         │
    Orchestrator (Opus)
    ┌────┼────┐
    │    │    │
 Worker Worker Worker
 (task) (task) (task)

Clawd: The OS Layer

[TODO: What makes Clawd different from project orchestrators]

Project Orchestrators

[TODO: Phillip, Oscar, etc. - domain ownership]

Parallel Execution with macx-tmux

[TODO: Link to /projects/macx-tmux, why tmux, the MCP pattern]

Model Routing

[TODO: Opus for orchestration, Haiku for exploration, Codex for shell work]

---

Previous: Daily Life With My Exocortex Next: Memory That Compounds