Technical Report · July 2026
The Room Protocol
Shared-State Coordination for Multi-Agent Software Development
Shreyas Padmakiran · roomd.sh
Abstract
LLM coding agents work well alone but have no built-in way to coordinate when several of them share one project. The usual workarounds (a human relaying state, agents chatting, or sharing only a repo) all treat coordination as message-passing. This paper argues for shared state instead: a small set of structured, persistent objects agents read and write.
The Room Protocol centers on one abstraction, the room. A room holds five primitives (plan, typed context, events, presence, locks) and is the unit of ownership. It is exposed over MCP so any capable client can join. The reference implementation, roomd, is a stateless server on one Redis store. The same choices that keep coordination durable also keep multi-tenant deploy simple.
roomd vs the Room Protocol
The protocol is the design; roomd is the server. Same split as HTTP versus nginx.
Contributions
- 01The room abstraction and five primitives for coordinating independent coding agents.
- 02Typed context: entries have a type and schema, so consumers read structure instead of parsing prose.
- 03A stateless reference server: fresh MCP handler per request over one Redis store.
- 04Multi-tenancy: team keys, first-touch room ownership, room-scoped invites, rate limits.
- 05Concurrency: plan locks plus per-agent event cursors.
- 06An evaluation plan for coordination overhead, handoff correctness, and concurrency safety.
Read the full report
The complete paper covers the protocol operations, the typed-context model, the stateless architecture, concurrency and access control, an evaluation methodology, and related work.