Agent-House Run
Most multi-agent workflows are hierarchical: one main agent calls subagents, chains, or workers. The orchestrator does all the thinking; workers are disposable. You lose peer-to-peer context, focused
Most multi-agent workflows are hierarchical: one main agent calls subagents, chains, or workers. The orchestrator does all the thinking; workers are disposable. You lose peer-to-peer context, focused windows, and the ability for agents to ask each other questions. A production agent with sensitive context cannot safely delegate to a debug agent without leaking data.
Agent House is a 7-room multi-agent workflow system where separate Pi sessions discover each other, send prompts, wait for replies, and coordinate as peers. Each room has an orchestrator + workers running in detached tmux sessions via Pi-to-Pi coms. Rooms: command-center (planning), intelligence (research), build-studio (implementation), quality-gate (review), launch-room (publish), growth-lab (metrics), product-design (UX). Results mirror into /root/agent-house/<room>/results/<timestamp>/. Handoff artifacts (HANDOFF.md) chain rooms together.
531 runs logged. RolePilot 3-room chain validated as end-to-end test. Workers coordinate via coms_list, coms_send, coms_await.
Active infrastructure used daily. 531 runs logged. Tested with a 3-room chain (product-design → build-studio → quality-gate) producing a validated onboarding prototype with 6 defects found, 6 fixed, 0 remaining. Next step: add automatic index/gallery generation so proof pages self-publish.