Technical Summary: AgentFugue

Problem Statement

Recent advancements in long-horizon agentic tasks have primarily relied on "scaling up"—improving individual agents through stronger foundation models, better tool integration, and more effective scaffolding. While this approach expands the capability of a single trajectory, it does not inherently increase the breadth of exploration.

The paper addresses the question of "scaling out": whether multiple peer agents working on the same task can collectively improve capability without relying on explicit role specialization, workflow orchestration, or centralized planning. The core challenge lies in communication: without it, agents operate as isolated searches; with unrestricted communication, useful signals are often overwhelmed by noise, and exploration diversity collapses. The authors argue that effective scaling out requires a mechanism for collective reasoning where peer agents can selectively exchange intermediate progress while maintaining distinct search directions.

Methodology: AgentFugue

The authors propose AgentFugue, a collective reasoning framework built around a shared reasoning hub. Unlike centralized planners or debate-style multi-agent systems, AgentFugue treats agents as peers that explore the same task in parallel, with the hub acting as an external communication layer rather than a controller.

Core Architecture

1. Shared Reasoning Hub ($H$): A plug-in communication layer that sits alongside peer agents. It does not replace local reasoning but makes intermediate discoveries from one trajectory selectively available to others.
2. Episodes: Local agent trajectories are divided into contiguous chunks called "episodes" based on a fixed context budget.
3. Two-Stage Hub Operations:
   • Episode Writing ($M_{write}$): When an agent's context reaches a write budget, the hub compresses the completed segment into a concise episode note ($z_{i,e}$). This note captures what was established, attempted, ruled out, and what evidence was collected. The raw episode content is stored in the hub, while the agent's working context is evicted and replaced by the note to free up capacity.
   • Intent-Driven Reading ($M_{read}$): Agents do not passively receive all notes. Instead, an agent issues a structured request containing an intent ($q_{i,t}$) and a set of episode references ($E_{i,t}$) based on the notes visible in its current context. The hub retrieves the full raw content of the referenced episodes and synthesizes a focused guidance or evidence piece ($r_{i,t}$) tailored to the agent's current need.

Optimization

The hub is instantiated using a moderate-sized language model (initialized from Qwen3.5-9B) and optimized in two stages:

1. Supervised Fine-Tuning (SFT): The hub learns to generate reference notes and readouts based on teacher-generated data.
2. Group Relative Policy Optimization (GRPO): The hub is further aligned with downstream task success via reinforcement learning within the full multi-agent loop. The task agents are frozen; the GRPO pressure optimizes the communication layer itself, rewarding hub outputs that lead to shorter, successful search paths.

Scaling Settings

The framework is evaluated in two settings:

• Homogeneous Teams: Multiple agents share the same backbone and configuration, testing gains purely from cross-trajectory interaction.
• Heterogeneous Teams: Agents differ in model backbones or configurations, testing whether the hub can mediate complementary reasoning biases and failure modes.

Key Contributions

1. Identification of Peer-Agent Scaling: The paper identifies a distinct setting for long-horizon reasoning where capability must arise from cross-trajectory reuse rather than role specialization.
2. AgentFugue Framework: A communication framework based on writing, retrieving, and reading shared reasoning episodes. It transforms parallel trajectories into a selectively shared reasoning ecology without centralized planning, drawing inspiration from the structure of a Baroque fugue where distinct voices develop one another's progress.
3. Empirical Analysis: A comprehensive study of the framework in both homogeneous and heterogeneous teams, analyzing when scaling out improves per-agent efficiency, when it yields team-level gains, and where the mechanism breaks down.

Experimental Results

The authors evaluated AgentFugue on three challenging long-horizon benchmarks: BrowseComp (deep multi-hop web search), WideSearch (breadth-oriented evidence collection), and HLE (expert-authored multi-domain reasoning).

• Performance: AgentFugue achieved state-of-the-art results across all benchmarks and backbone configurations (Qwen3.5-35B-A3B and DeepSeek-v4-Flash), outperforming strong baselines including single-agent ReAct, DeepResearch systems, and other multi-agent approaches (Naive-Multi-Agent, Swarm-Multi-Agent).
• Scaling Behavior:
  • In homogeneous teams, per-agent accuracy improved significantly as team size increased from $N=1$ to $N=5$, after which it plateaued. However, team-level aggregation (Pass@N) continued to improve with larger teams.
  • In heterogeneous teams, weaker models benefited the most from the hub, while stronger models also saw gains. Heterogeneous teams triggered higher hub usage (memory traffic) compared to homogeneous ones, indicating that diverse backbones produce intermediate notes that are genuinely new to peers.
• Efficiency: The hub successfully traded isolated exploration for shared coordination. As team size grew, per-agent search and visit calls decreased, while memory traffic increased, indicating a shift from redundant exploration to structured information reuse.

Significance and Claims

The paper claims that collective reasoning can turn the scaling of peer agent systems into a distinct source of capability gains, rather than merely a method of spending more compute.

• Distinct from Existing Paradigms: Unlike multi-agent debate or group chat, AgentFugue does not force synchronized turn-taking. Unlike "best-of-N" sampling, trajectories influence one another during exploration via reusable intermediate progress. Unlike RAG over static corpora, the read path is intent-driven and synthesizes raw content on demand.
• Generalization: The improvements were consistent across retrieval-heavy, reasoning-centric, and breadth-oriented tasks, suggesting the shared reasoning hub is a generic coordination primitive.
• Limitations: The authors note that the current implementation uses a moderate-sized hub and a limited set of backbones. They also acknowledge failure modes where low-quality notes or overconfident intermediate conclusions can propagate misleading hypotheses across the team, potentially reducing diversity.

The authors conclude that while the challenge remains to make such communication selective enough to preserve diversity while reliable enough to prevent error propagation, AgentFugue demonstrates that peer-agent scaling is a viable and effective axis for improving long-horizon agentic reasoning.