Researchers from Coral AI Labs and several universities have released AgentRadio, an open-source messaging layer that lets coding agents share findings mid-task without stopping their own work. The project, published on arXiv, targets a known weakness of multi-agent systems: most designs force agents to either work or communicate, but not both.
AgentRadio gives agents three primitives: opening a conversation thread, appending messages without blocking, and waiting for a mention while continuing to run. The system prompt tells each agent to keep a background watcher active, so a peer’s discovery lands without interrupting the main task. The authors describe the result as passive awareness, and it requires no changes to the underlying model or harness.
On 124 long-horizon questions from the SWE-Atlas QnA benchmark over production repositories, a single Claude Code instance on Claude Opus 4.6 solved 32.3 percent, and upgrading to Opus 4.8 only reached 57.2 percent. A team of four agents using AgentRadio nearly doubled the accuracy of four independently working agents and outmatched single agents running more advanced models.
The research suggests that coordination structure can beat raw model scale, since interdependent subtasks like codebase investigation benefit from mid-course corrections instead of waiting for a formal review phase. AgentRadio is Apache 2.0 licensed and plugs into existing harnesses like Claude Code or Codex CLI through three shell scripts.
