Coral AI Labs develops AgentRadio for real-time AI agent coordination

Getting AI agents to work together on complex tasks has always had an awkward problem. They can divide the work just fine. They just can’t talk to each other while they’re doing it.

Coral AI Labs, working alongside researchers at multiple universities, published a paper on July 30 introducing AgentRadio, an asynchronous message-passing layer that lets AI agents communicate between execution steps without pausing their primary workflows.

The coordination gap in multi-agent systems

Most multi-agent AI architectures force agents to communicate at rigid checkpoints, like the end of a phase or during synchronized rounds. That works well enough for simple, modular tasks where each agent’s work is neatly self-contained.

AgentRadio addresses this by creating what the researchers call “passive awareness.” Agents can receive and process insights from their counterparts without stopping what they’re doing. The architecture is built on three core primitives: create_thread, send_message, and wait_for_mention.

In empirical tests using SWE-Atlas QnA, a benchmark for complex code analysis, a team of four Claude Code agents equipped with AgentRadio resolved 62.1% of tasks. A single Claude Code agent, working alone, managed just 32.3%. That’s a 29.8 percentage point improvement.

Beating even stronger baselines

The more interesting comparison is against Opus 4.8, a more powerful single-agent baseline. Even Opus 4.8 only achieved a 57.2% resolution rate on the same benchmark. Four coordinating agents using a less powerful model outperformed a single agent using a stronger one.

The gains were most pronounced on harder queries, exactly the kind of long-horizon tasks that enterprise teams actually care about.

The paper, available on arXiv (2607.28430v1), details how the asynchronous design avoids the overhead that typically comes with multi-agent coordination. Agents don’t block while waiting for messages. They check for relevant updates opportunistically, processing new information when it’s useful and ignoring it when it’s not.

Open source and broader infrastructure

Coral AI Labs released AgentRadio as open-source software under the Coral-Protocol organization on GitHub. The release fits into a broader ecosystem the organization is building, including CoralOS, a system designed for orchestrating AI agents at scale.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.