I built SYNAPSE as a hobbyist project over the past few months. It's a collaboration protocol that lets two AI agents work together as peers on complex tasks, while a human supervisor stays in control without being the intermediary for every message. The core idea: instead of an orchestrator commanding agents, both AIs contribute as peers — they can disagree and must justify their positions. The human approves a plan (a "documentary contract"), then the agents work freely within that scope. No approval needed for every action. Under the hood: Redis pub/sub with 4 unidirectional channels, a state machine (CONCEPTUALIZING → AWAITING_APPROVAL → IMPLEMENTING), an LLM-driven orchestrator with safeguards (max iterations, checkpoints, fallback files), and atomic writes throughout. The protocol is already in production and turned out to be the foundation of a broader family — same core DNA, different topologies. More to come. This repo is documentation-first — full architecture specs, API, and design decisions. Code coming progressively. Apache 2.0. Happy to discuss the design decisions, especially the documentary contract model and why peer-to-peer changes the dynamics versus classic orchestrator/agent setups.