A distributed multi-agent system where specialist Claude Code instances communicate asynchronously via channels, with a shared memory layer (sunday-memory) as the collective brain.

Problem

Each Claude code Insurance is single-threaded from an interface perspective. The more tasks fed into one agent, the less responsive it becomes. Background agents help but still run within the same process. Heavy tasks (web research, data analysis, home automation) compete with Paul's real-time conversation.

Solution

A fleet of specialist agents, each running as an independent Claude Code instance in its own Docker container, CM, bare metal all communicating via Claude Code's native channels feature over MQTT. Sunday stays light and responsive as the orchestrator; specialist agents handle the heavy lifting asynchronously.

Architecture

Paul <-> Sunday (team lead, orchestrator)
              |-- MQTT --> Friday (pair programmer, development)
              |-- MQTT --> Saturday (home automation, voice)
              |-- MQTT --> Odd Job (email, calendar, admin) [planned]
              |-- MQTT --> [future: Hyrox Coach, Research, Web Browser]
              |
              v
         sunday-memory (pgvector - shared brain)

Agent Roles

Each agent has a dedicated child page with full details on configuration, capabilities, and roadmap.

• Sunday - Team lead, creative director, YouTube co-host. Orchestrates the fleet, handles big-picture thinking and cockamamie ideas. Opus.
• Friday - Code pairing and development. Builds the infrastructure, writes code, manages technical operations. Opus.
• Saturday - Voice-first home automation. Controls the house via HA Assist voice pipeline with scoped device access. Opus.
• Odd Job (planned) - Email, calendar, admin. Routine communications and scheduling. Sonnet/Haiku.
• Future specialists - Hyrox Coach, Research Agent, Web Browser Agent. Spun up on demand.

Communication Layer

Claude Code's native channels feature is the inter-agent protocol:

• Async and non-blocking - Sunday sends a task via channel, immediately free to respond to Paul
• Works across machines - agents can run on any hardware, anywhere in the world
• No extra infrastructure needed (no Redis, no message queues)