The paper "AgentRadio: Passive Awareness for Long-Horizon Multi-Agent Collaboration" introduces a novel asynchronous communication layer designed to improve the performance of Large Language Model (LLM) coding agents in complex, long-horizon tasks such as codebase comprehension. The authors address the limitation where current multi-agent systems either work in isolation or suffer from synchronized "blocking" communication, which forces agents to halt their productive execution to share information or wait for team updates.

Core Methodology: AgentRadio

AgentRadio equips agent harnesses with a message-passing layer defined by three specific primitives:

1. create_thread(name, participants): Initializes a named conversation on a shared message server.
2. send_message(thread, content, mentions): An immediate, non-blocking operation that broadcasts data or targets specific teammates.
3. wait_for_mention(timeout): The critical innovation. When run as a background task, it allows an agent to remain "passively aware" of its peers. Messages surface between the agent's foreground work steps without interrupting its active command execution.

Mathematically, let an agent's execution be a sequence of steps $s_1, s_2, \dots$ and $M(t)$ be the set of visible teammate messages at step $s_t$. In a traditional blocking system, $M$ only grows if the agent spends a step calling wait_for_mention. In AgentRadio, $M(t)$ accumulates all messages sent before $s_t$ without consuming any foreground steps, decoupling communication from the agent's primary work loop.

The Five-Phase Protocol

The authors evaluate this layer using a standardized protocol of division of labor and negotiation:

• P1, Explore: Independent repository exploration.
• P2, Divide: Assembler (agent-1) coordinates sub-question partitioning through a planning thread.
• P3, Execute: Concurrent sub-task execution. Under AgentRadio, discoveries (e.g., dead ends, contradictory findings) are posted to a worklog in real-time, allowing teammates to fold this information into their ongoing work immediately.
• P4, Review: Cross-examination of evidence where conflicts or thin claims can trigger a return to P3.
• P5, Submit: Final synthesis of the collaborative output.

Performance and Analysis

Testing on the SWE-Atlas QnA benchmark (124 tasks) using Claude Code and DeepSeek V4 Pro, the authors demonstrate that AgentRadio significantly outperforms both single-agent baselines and alternative multi-agent configurations. Specifically, four Claude Code agents using AgentRadio achieved a 62.1% resolution rate, exceeding the 32.3% of a single agent and the 57.2% achieved by the advanced Claude Code Opus 4.8.

The paper provides a detailed rubric-level analysis showing that:

• Passive awareness provides a modular gain: The "passive step" alone improved task accuracy by 10.5–11.3 percentage points over the blocking negotiation protocol.
• Difficulty correlation: The benefit of passive awareness increases with task difficulty, particularly in architectural and system-design questions where cross-boundary information is vital.
• Structure over compute: The gains are attributed to the coordination structure rather than increased compute budget, as "best-of-6" sampling configurations failed to match the performance of the full AgentRadio stack.

The paper concludes that for tasks involving interdependent sub-problems, the ability to maintain passive awareness of team discoveries—without forcing agents to pause their internal reasoning—is a critical factor in enabling successful, long-horizon collaboration.