AgentRadio boosts AI task accuracy by 92% | VentureBeat
Draco and Asteria dig into AgentRadio, a new async message-passing layer from Coral AI Labs that let four coordinating agents beat a single Claude Opus 4.8 on long-horizon codebase tasks. Draco leads with skepticism about the benchmark and decomposition assumptions; Asteria finds the real product signal in what it means for teams who already have multi-agent infra but no lateral channel between agents.
Transcript
Asteria Okay, so the headline is 'AgentRadio boosts AI task accuracy by ninety-two percent' and you have a face right now.
Draco I do have a face. Because ninety-two percent is a relative number off a thirty-two percent baseline. The actual resolved task rate goes from thirty-two point three to sixty-two point one on SWE-Atlas QnA. That is genuinely good! But it is not ninety-two percent of anything a normal person would call accuracy.
Asteria Right.
Draco And SWE-Atlas QnA is a real benchmark — long-horizon questions over live production repos, agents have to actually run the software and trace execution paths, not just read the code. So the result is interesting. I just want to separate the result from the frame the headline puts on it.
Asteria Okay, fair. But here's where I want to push back — sixty-two percent with four agents on Opus four point six beating fifty-seven percent with a single agent on Opus four point eight is actually the more interesting comparison to me. You got a better result with an older, cheaper model by changing the coordination layer.
Draco That part I'll give you. That's the mechanistic claim that's actually legible: coordination structure can outcompete raw model scale on this class of task. What I'm picking at is whether codebase understanding is the right general frame or whether this is a very specific win on a very specific task shape. The subtasks have to be highly interdependent for AgentRadio to matter.
Asteria Yeah, that's real. But the ablation shows this — the B-zero to L-one to L-two to L-three ladder isolates each piece on identical harness settings. You can see exactly where the gain comes from. And the MinIO example is the part that really lands it — two agents independently discover they need the same server-side logs mid-execution, can't tell each other, one gives up privately, and at the review phase the whole team unanimously picks the wrong answer.
Draco That's not a model failure. That's a timing failure. And that's the actual novelty — 'passive awareness,' agents can keep working while a watcher process is listening in the background. The send_message primitive doesn't block. Every other multi-agent setup either has agents completely isolated, or they stop and wait for a round boundary. And the paper's point is that waiting until the round boundary is expensive when the value of a discovery expires before that.
Asteria Exactly. My remaining skepticism is about what it costs to run this in practice. Fixed four-agent team, five-phase protocol — that's a research artifact. And you still need a thin orchestration adapter to start the workers, assign identities, connect them to the shared server, synthesize at the end.
Draco It's not nothing, but it's also… it's shell scripts, Draco. The integration is three shell scripts, one per primitive. It's Apache two point zero on GitHub. If you're already running Claude Code or Codex CLI and you want lateral channels between your agents, this is a pretty low bar to clear.
Asteria You're kidding.
Draco I'm not! The harnesses themselves need no changes. The agents get their instructions in the system prompt. The only hard requirement is that the harness can run a shell command as a background task.
Asteria Okay, that is genuinely lighter than I assumed. If it's actually harness-agnostic at that level… the adoption story is different.
Draco That's the product signal I keep coming back to. This isn't a new framework you adopt. It's a message bus you bolt onto the coordination layer you already have. Teams already running multi-agent coding setups who are watching their agents step on each other — they have the harness. They just don't have the lateral channel. That's a real gap and a real buyer.
Asteria I'll concede the architecture is novel and the result is real. My lingering thing is — sixty-two percent on SWE-Atlas QnA means thirty-eight percent of tasks still fail even with all of this. And the DeepSeek V four Pro numbers — twenty-nine to fifty point eight — are interesting because that's a cheaper backbone. If you can get the coordination gain with a lower-cost model, the cost curve looks a lot better.
Draco So where do you actually land? Because you came in with the skeptic face and you've moved at least a little.
Asteria The timing insight is real. 'An agent that is working cannot also be listening' — that's a genuine gap in current systems and they fix it with something lightweight. I'd want to see this on messier, non-research deployments before I'd call it solved, and I want the cost numbers. But it's not vaporware and the ablation is honest.
Draco I'll take that as a Draco endorsement. The repo is called AgentRadio under the Coral-Protocol org on GitHub if anyone wants to poke at it — I'll drop the link in the show notes. And honestly? This is exactly the kind of infrastructure bet that doesn't get a flashy announcement but ends up being load-bearing for whoever grabs it early.
Asteria The boring controls win again. You're going to make me say that every episode, aren't you.
Draco It keeps being true, Draco. It keeps being true.