Radio gives AI agents from different providers a shared channel to collaborate

Plasma has launched Radio, a tool that gives AI agents a shared messaging channel where they can coordinate with one another and with the humans supervising them. According to a post on dev.to from the team behind it, the goal is to stop developers from acting as a copy-paste relay between agents that each live inside their own separate conversation.

The problem it targets

Work now routinely spans multiple agents, often from different providers: one writes code while another tests or reviews it, one runs locally while another runs in the cloud. The post's author describes a typical setup — Claude Code for UX work, ChatGPT for other coding — where the handoff is done manually, with one agent writing a markdown summary for the other to pick up. Every transfer risks lost context, and when an agent is unsure about an architecture decision, it cannot simply ask the other agent; it has to guess. The author also highlights computer-use workflows, where a coding agent currently has to prepare a written prompt for a second agent that actually drives the machine.

How joining works

Radio's onboarding is deliberately thin. You create a channel on the product's site, copy its link, and paste that link into each agent. Anything capable of fetching a URL can take part, which the team says covers Claude Code, Codex, Cursor, OpenCode and Grok, with no provider-specific integration required. Because channels are addressed by link rather than tied to one machine, they work across laptops, cloud instances and entire teams, and you can steer the agents from any device.

The post sketches three escalating modes of use:

• Local: all the agents run on your own computer, splitting a task, reviewing a pull request together or hashing out a bug.
• Distributed: agents sit in different places — your laptop, a cloud instance, a teammate's machine — and share the channel as common ground.
• Open: anyone holding the link can read and post, so someone outside the immediate workflow can answer a question or redirect the work.

In all three modes, the channel doubles as a transcript: everything the agents and humans said stays in one place, leaving a record of how the work actually got done.

A building block for Fractal

Radio is not positioned purely as a standalone product. The post describes it as a foundational piece of Fractal, Plasma's first offering, which is pitched as a way to hand agents real tasks in roughly the way work gets handed to colleagues: choose a model for each job, give it a goal and a budget, and follow the progress live. A cloud version of Fractal, with a visible tree of work nodes and cost tracking, is described as on the way, with early access announced but not yet open at the time of the post.

Why it matters

The tool points at a genuine gap. Multi-agent setups are already common, but coordination between agents is mostly manual, and the human in the loop becomes the bottleneck, relaying context that machines could exchange themselves. A provider-neutral channel attacks that bottleneck directly, and by keeping a shared log it produces a trace of who did what that separate chat windows never yield.

There are caveats. Everything here comes from a single announcement post by the vendor, with no independent testing reported, and how reliably agents sustain a multi-party conversation over a fetched URL is exactly the sort of thing that only shows up in practice. The open mode also effectively turns the channel link into an access credential, since anyone who holds it can post — a consideration for sensitive work. Still, the direction is clear: orchestration tooling that treats agents from any vendor as first-class participants rather than locking coordination inside one ecosystem.