Agent Teams Beat Claude Opus on Coding Tasks

Four Claude Code agents working as a coordinated team nearly doubled their task accuracy on enterprise coding benchmarks compared to the same four agents working alone, according to researchers at Coral AI Labs and several university collaborators. The team also beat single agents running on more advanced models, including Claude Opus. The gain came from a new communication layer called AgentRadio, which lets agents pass messages to each other mid-task instead of waiting for a scheduled check-in.

That distinction matters more than it sounds. Most multi-agent systems today split work among agents but only let them compare notes after each one finishes its assigned chunk. For enterprise codebases, where one function's behavior often depends on decisions made three files away, that lag is expensive.

An agent can spend hours tracing a dead-end path before anyone catches the mistake. AgentRadio's asynchronous message-passing lets agents flag problems, share findings, and redirect each other's work in real time, without stopping to finish a formal review cycle first.

The researchers tested this against a benchmark of long-horizon questions over production repositories, the kind of sprawling, interdependent code that breaks single-agent systems in a specific, well-documented way.