The Big Picture
Background listening between working coding agents—so messages surface between steps without interrupting work—nearly doubles task success on long software-understanding problems, while requiring only a small harness change.
ON THIS PAGE
Key Findings
Adding a lightweight message layer where agents can listen in the background (passive awareness) lets teammates’ discoveries be folded into ongoing work without stopping the agent. Four coding agents using that layer solved 62.1% of long, real-world code questions versus 32.3% for a single agent, and the passive-listening feature itself contributed a large, statistically significant chunk of the gain. The approach beats compute-matched sampling strategies and helps most on harder tasks where sequential context and mid-task discoveries matter. passive-listening feature
Data Highlights
162.1% task accuracy with four Claude Code agents plus AgentRadio (Opus 4.6) versus 32.3% for one agent — a +29.8 point (92% relative) gain.
2Passive awareness (background listening) alone adds +10.5 points with Opus 4.6 and +11.3 points with DeepSeek V4 Pro in paired tests.
3Full AgentRadio protocol outperforms compute-matched best-of-6 sampling: 62.1% vs 37.9% (Opus 4.6) and 50.8% vs 31.4% (DeepSeek V4 Pro).
What This Means
Engineers building multi-agent systems and orchestration layers: a small change to when agents listen can dramatically improve long, multi-step tasks like codebase understanding. Technical leaders evaluating agent fleets: you can gain reliability without swapping models—change the harness, not the model. Researchers: this is a practical, measurable way to study asynchronous coordination and agent-to-agent communication. asynchronous coordination
Avoid common pitfallsLearn what failures to watch for
Key Figures
Fig 1: Figure 1: Task accuracy on SWE-Atlas QnA (124 tasks). Four Claude Code agents organized by AgentRadio surpass the single agent of both models and the strongest single-agent entry on the public leaderboard, Claude Code with the newer Opus 4.8.
Fig 2: Figure 2: AgentRadio and the five-phase protocol. Left: each Claude Code agent runs its harness work in the foreground while exactly one background watcher waits for mentions, so incoming messages surface between work steps and never interrupt a running command. Sending is non-blocking, and all traffic flows through named threads on a shared message server. Right: the five phases with four agent lanes. During Phase 3 an agent posts a discovery at the moment it is made, the teammate’s lane never breaks, and the message is folded into the ongoing subtask at the next step. The message shown is illustrative.
Fig 3: Figure 3: The task mix of SWE-Atlas QnA: 124 questions over 11 production repositories in four languages (Raghavendra et al. 2026 ) .
Fig 4: Figure 4: Rubric-level attribution of the Opus 4.6 ladder, as a waterfall from the single agent to the full protocol. Each floating bar gives the net change of one layer, with the gross rubrics gained and lost below it. The dotted line marks the benchmark total of 1,306 rubrics.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Results come from a benchmark of long code-understanding questions (124 tasks on SWE-Atlas QnA) and assume agents can build, run, and trace software in a sandbox; different domains may behave differently. The benefit depends on a five-phase protocol (division, negotiation, review) and a harness that supports non-blocking sends and background waiting—without those pieces the gain may shrink. Messages can sometimes distract agents mid-evidence-gathering and cause regressions on some rubrics, so monitoring and message design matter. non-blocking sends
Deep Dive
AgentRadio is a simple, asynchronous message layer that gives each coding agent three operations: create a conversation thread, send a message without blocking, and wait for mentions. The key idea is running the wait-for-mention operation as a background task so agents keep executing their normal steps while mentions from teammates are recorded and surfaced at the next step boundary. That passive awareness removes the usual trade-off where an agent must stop working to receive messages. In controlled experiments on 124 challenging code-understanding questions, the authors compare single-agent baselines, repeated single-run sampling, and multi-agent ladders where teams divide work, negotiate plans, and cross-review results. Switching only the listening mode—from blocking receives to passive background listening—yields large, paired gains: with four agents the full protocol reaches 62.1% task accuracy (Opus 4.6) and 50.8% (DeepSeek V4 Pro), substantially above single-agent and best-of-6 sampling baselines. Rubric-level analysis shows division of labor expands coverage, negotiation reduces harmful fragmentation, and passive awareness recovers mid-execution discoveries that would otherwise be lost. The practical takeaway: a small harness change—letting agents hear without stopping—gives outsized returns on long-horizon, interdependent tasks. multi-agent ladders A2A Protocol Pattern
Need expert guidance?We can help implement this
Credibility Assessment:
All authors have low h-index values and no affiliations or citations are listed; venue is an arXiv preprint. Signals point to emerging or limited-information contributors.