AI agents beat Claude Opus 4.8 in coding tasks

Four AI coding agents, working together in real time, just beat one top model on tough enterprise software tasks. In tests run by Coral AI Labs and university partners, a team of four Claude Code agents coordinated via a new system called AgentRadio and solved complex coding problems more accurately than a single state-of-the-art model, Claude Opus 4.8. The breakthrough happened on live enterprise codebases, where tasks often span hundreds of files and require dozens of tool calls.

Most AI agents today handle long jobs by breaking them into steps and passing the baton, but they rarely talk while working. That silence leads to dead ends and wasted time. AgentRadio fixes this by letting agents chat between steps—without pausing their work. It acts like a background radio channel, letting agents swap small updates, flag issues, and reroute tasks on the fly. For sprawling codebases, where a change in one file might break another, this real-time feedback loop cuts down on retries and speeds fixes.

In benchmarks on production repositories, the team of four agents cut errors nearly in half compared to a lone high-end model running the same jobs. Researchers measured both correctness and completion time, finding that the coordinated squad finished tasks faster and with fewer follow-up corrections. Early users at two Fortune 500 companies reported fewer rollbacks and quicker deployments when swapping their single-agent setups for the multi-agent system.

The next step is scaling AgentRadio beyond coding into broader enterprise workflows like security reviews and compliance checks. If the pattern holds, teams may soon deploy fleets of specialized agents that talk to each other instead of waiting for human oversight. That shift could slash the time and cost of maintaining large software systems while raising the bar for AI reliability in real-world environments.