AgentRadio's passive awareness lifts multi-agent coding accuracy to 62.1%

Long-horizon coding tasks—like understanding a production codebase—require agents to build software, trace execution, and synthesize evidence across files, often over tens of minutes. On the SWE-Atlas QnA benchmark, a single Claude Code agent (Opus 4.6) resolves only 32.3% of tasks. Dividing the work among multiple agents with clean contexts helps, but subtasks are highly interdependent: one agent's discovery can completely rewrite another's task. Existing multi-agent systems only synchronize at phase boundaries, so findings made mid-execution sit idle until a handoff or round completes. Communication and work remain mutually exclusive, limiting coordination on the fly.

AgentRadio removes that bottleneck. Developed by Xinxing Ren and colleagues, it equips coding-agent harnesses with three primitives: threads, messages, and waiting for mentions. The last one runs as a background task, surfacing teammates' messages without interrupting foreground work—so each agent stays passively aware of peers and can fold new findings into its ongoing task. Under a five-phase protocol of division of labor and negotiation, four agents using AgentRadio resolved 62.1% of tasks on SWE-Atlas QnA—29.8 points higher than a single agent and above Claude Code with the newer Opus 4.8 (57.2%). Rubric-level analysis shows the gain grows with task difficulty, suggesting mid-course correction is the key mechanism. The code is publicly available, giving multi-agent systems a way to share discoveries in real time without stalling execution.