Four AI coding agents, each running a merely capable model, were given a background channel to overhear each other mid-task - no meetings, no interruptions, no extra spend. The team beat a single agent running a newer, stronger model by a wide margin, and a compute-matched version of "just try more times" by nearly double. The lesson isn't about smarter models. It's about what a team loses when insight has to wait for a scheduled sync.
In plain English
In plain English: Letting AI agents working on a shared problem overhear each other's discoveries in real time, without interrupting anyone, makes the whole team dramatically better at hard, interdependent tasks.
Like this: Four coworkers on a hard investigation, each digging into a different thread. Normally they only compare notes at the 3pm meeting - so a 10am discovery just sits there. This is the walkie-talkie channel that's always on: nobody stops working, but the moment someone finds something, it reaches the rest of the team instantly.
Example: Four "regular" agents, coordinating well, outscored one "genius" agent working alone - proof that how a team communicates can matter more than how capable any single member is.
The Always-On Channel Nobody Has To Check
The system is called AgentRadio, and it's built from three primitives: open a thread, send a message, wait for a mention. The critical design choice is that "waiting" runs as a silent background process - a teammate's message surfaces between an agent's own work steps, never interrupting a command it's mid-way through running. No extra model calls are spent checking for messages. It's just an ordinary background process, bolted onto the outside of the agent harness with zero changes to how any individual agent reasons.
124 Questions One Agent Alone Mostly Failed
The test was genuinely hard: 124 long, open-ended questions about real production codebases, where answering requires actually building and running the software, tracing behavior across files - not reading and guessing. A single AI agent resolves only about a third of these on its own.
Four agents working under a simple five-step process - explore, divide the work, execute, review, submit - hit 62.1% when they could overhear each other in real time. That beats even a single agent running a newer, stronger model at 57.2%, the best solo score on the public leaderboard at the time.
It's Not Just More Compute
The obvious objection: isn't this just four times the resources? The researchers checked directly, running a single agent six separate times on the same problems and keeping its best attempt - spending almost the exact same money as the four-agent team.
Same rough compute budget, very different structure
Same rough spend as the best-of-six attempt. Nearly double the score. The coordination structure, not the extra money, explains almost the entire gain.
"A great communication system doesn't make a team smarter. It makes a team lose less of what it already knows."
The Discovery That Almost Died Silently
There's a case study in the paper worth sitting with. The team was debugging a storage server issue. One agent quietly found the exact fix - a specific configuration switch - but, working heads-down, never got around to formally announcing it. Under the old handoff-only setup, that discovery died silently. The team submitted the wrong answer.
Under real-time overhearing, the same discovery reached the rest of the team the instant it happened. It flipped five wrong answers into right ones.
The Most Honest Result In The Paper Is A Zero
Not every failure got fixed by better communication, and the paper doesn't pretend otherwise. In a second case study, the team needed a conclusion that none of the four agents had privately reached. Overhearing each other added exactly zero benefit here - there was nothing to overhear. The system shares what someone already figured out. It can't invent an insight nobody had.
That's the sharper, more useful way to think about what any real-time-coordination tool actually buys a team: not raw intelligence, but retention. It stops a team from quietly losing the one thing that would have made the difference. And the design choice to never touch how any individual agent reasons - the entire coordination layer bolts on from the outside, as background shell scripts - is what makes it something any team already using coding agents could plausibly add tomorrow, not a research prototype requiring a rebuild.
