Researchers present AgentRoom, a real-time collaborative editing protocol that enables concurrent multi-agent LLM coding through a shared filesystem built on Conflict-free Replicated Data Types, exposing file-level operations as MCP tools. Evaluated on five frontier models across backend coding tasks in Python and Rust, the system shows that coordination mechanisms, rather than parallelism alone, drive performance improvements. Two-agent AgentRoom configurations abandon fewer tasks than solo baselines with less run-to-run variation, with gains stemming primarily from coordination rather than CRDT merging or raw parallelism.
