Coding Agents That Message Mid-Task Nearly Double Their Accuracy

AgentRadio lets four coding agents swap messages mid-task, nearly doubling their score on a codebase benchmark, per a paper posted to arXiv.

That's the headline number for anyone weighing multi-agent coding tools. Four agents talking scored 62.1% on the SWE-Atlas QnA benchmark, versus 57.2% for a single agent running the newer Opus 4.8 model alone.

The system runs on three primitives: threads, messages, and waiting for mentions. That wait step runs as a background task, so an agent stays aware of teammates without blocking its own work.

A single Claude Code agent on Opus 4.6 answers 32.3% of these long-horizon codebase questions alone. The four-agent gain grows with question difficulty, which the paper attributes to mid-course correction rather than raw parallelism.

The paper doesn't say what model the four-agent team itself ran on, only that it beat a solo run on Opus 4.8. That gap is the real test for teams deciding where to spend engineering time: on agent coordination, or on the next model release.

The framework's code is public, released under Coral-Protocol.