AgentRadio lifts four-agent accuracy to 62.1%

AgentRadio’s four Claude Code agents resolved 62.1% of 124 SWE-Atlas QnA tasks, versus 32.3% for one Opus 4.6 agent and 57.2% for one Opus 4.8 agent, after researchers gave peers a background channel for sharing discoveries mid-run. The open paper and implementation also put the coordination tax on the record: average spend rose from $2.96 to $19.45 per task, while six independent runs costing $17.76 reached only 37.9%. For teams extending the evidence that coding-agent harnesses can swing results, the operator implication is to reserve live multi-agent coordination for interdependent, high-cost investigations and benchmark cost per accepted answer—not agent count or token price.