MonoScale: Scaling Multi-Agent System with Monotonic Improvement

TL;DR

MonoScale tackles open-ended MAS expansion where adding new agents risks cold-start misrouting and performance collapse. It introduces expansion-aware familiarization and auditable memory updates, framed as a contextual bandit with a trust-region surrogate to ensure monotonic improvement across onboarding rounds, even as the action space grows. The memory distills successes and failures into routing principles that guide future decisions, enabling stable gains as the agent pool scales from 3 to 10 and improving robustness to noisy agents. Theoretical results prove a stage-wise monotonic non-degradation guarantee, and experiments on GAIA and Humanity's Last Exam show competitive performance with smaller backbones and resilience beyond static baselines, highlighting practical impact for scalable, open MAS deployments.

Abstract

In recent years, LLM-based multi-agent systems (MAS) have advanced rapidly, using a router to decompose tasks and delegate subtasks to specialized agents. A natural way to expand capability is to scale up the agent pool by continually integrating new functional agents or tool interfaces, but naive expansion can trigger performance collapse when the router cold-starts on newly added, heterogeneous, and unreliable agents. We propose MonoScale, an expansion-aware update framework that proactively generates a small set of agent-conditioned familiarization tasks, harvests evidence from both successful and failed interactions, and distills it into auditable natural-language memory to guide future routing. We formalize sequential augmentation as a contextual bandit and perform trust-region memory updates, yielding a monotonic non-decreasing performance guarantee across onboarding rounds. Experiments on GAIA and Humanity's Last Exam show stable gains as the agent pool grows, outperforming naive scale-up and strong-router fixed-pool baselines.

Figures (4)

• Figure 1: Overview of our expansion-aware familiarization-and-memory-update protocol. After adding a new agent, we generate customized warm-up tasks (conditioned on agent cards), collect both success and failure traces under the current router, distill the evidence into structured routing principles (memory candidates), and select a safe memory update with a conservative fallback to prevent brittle mis-routing during scale-up.
• Figure 2: Scaling performance on GAIA as the agent pool grows from $N=3$ to $N=10$. MonoScale consistently improves with more agents by updating routing memory during expansion, mitigating the performance collapse that can occur under naive scale-up.
• Figure 3: GAIA performance as the Malfunctioning Agent Pool scales from 3 to 10: a level-wise (L1–L3) and overall accuracy comparison. Gemini-3-Pro without memory updates exhibits a clear performance collapse during scaling, while our Qwen3-30B-based MonoScale remains stable and even improves slightly as the pool grows.
• Figure 4: Additional results on the noisy/malfunctioning agent pool in GAIA. We report level-wise (L1--L3) and overall accuracy as the agent pool scales ($N=3,5,7,10$) for multiple routers and inference variants, including a best-of-$3$ aggregation (pass@3) for the Qwen3 (w/o memory) router.

Theorems & Definitions (6)

• Lemma 4.1: Exact Bandit Surrogate
• proof
• Lemma 4.2: Expansion Bridge
• proof
• Theorem 4.3: Monotonicity Across Expansions
• proof