Concurrent multi‑agent coding can divide labor across modules, improve robustness through redundancy, and explore parallelly at the natural granularity of multi‑file projects. Real‑time collaborative editing protocols solve the coordination problem for human teams via Conflict‑free Replicated Data Types (CRDTs), but underlying large language models (LLMs) generate one token at a time, and existing multi‑agent coding systems inherit this serial limitation: they either sequence agents through phase handoffs or pool independent samples without coordination, and a single agent abandons roughly half of hard tasks when faced with a one‑file stub‑and‑exit scenario. AgentRoom introduces a real‑time collaborative editing protocol for concurrent coding agents. Its runtime layer exposes file‑level claim, status, and broadcast as MCP tools on a CRDT‑merged shared filesystem. We evaluated five frontier coding‑CLI models on four backend coding tasks, with cross‑language checks using Python DevBench and Rust+axum. For CLI‑stable models, AgentRoom with two agents abandons fewer tasks than solo execution and exhibits less run‑to‑run variation. At matched compute, a positive LLM‑judge contrast places AgentRoom above parallel‑merge, while a bundle probe contrast shows the full AgentRoom surpassing each partial case: it provides an ordering rather than a percentage split. The results indicate that coordination, not mere parallelism or CRDT‑merge, bears the performance load.
Blogger's Review: AgentRoom demonstrates how adding a real‑time collaborative layer can unlock the potential of multi‑agent systems, especially for complex coding tasks that span multiple files. By leveraging CRDTs for consistency and file‑level claim/broadcast mechanisms for effective task allocation, it markedly reduces task abandonment, warranting further exploration in larger‑scale coding projects.