Four Claude Code agents running on Opus 4.6 and coordinated by AgentRadio solved 61.4% of long-horizon enterprise coding tasks on the SWE-Atlas QnA benchmark, beating a single agent on the more advanced Opus 4.8 model, which scored 57.2%. A single Opus 4.6 agent managed only 32.3%. The system, built by researchers at Coral AI Labs and multiple universities, proves that coordination architecture can outperform raw model scale.
The core problem AgentRadio solves is mutual exclusion: in existing multi-agent systems, an agent that is working cannot also be listening. Current architectures either run agents in parallel with no communication, synchronize them at round boundaries forcing costly waits, or offer only top-down task dispatching with no lateral peer-to-peer channels. AgentRadio inserts a lightweight asynchronous message-passing layer between execution steps using three primitives: create_thread, send_message, and wait_for_mention. The wait_for_mention script runs as a background process, delivering a full thread snapshot the moment another agent mentions the caller, without interrupting the primary task. No changes to the underlying agent harness are required, only Apache 2.0 licensed shell scripts and a standalone message server.
The paper is worth reading in full for its breakdown of why codebase understanding is uniquely hostile to clean task decomposition. A single configuration file or mid-investigation bug discovered by one agent can invalidate another agent's entire hypothesis, and round-synchronized systems let that bad hypothesis run to completion before correction is possible. The researchers tested 124 tasks across system design, root-cause analysis, security, and API integration using both Claude Opus 4.6 and DeepSeek V4 Pro. The technical architecture section, covering harness-side integration and the passive awareness state, is the part practitioners will return to when building their own implementations.
[READ ORIGINAL →]