MonoScale: Scaling Multi-Agent System with Monotonic Improvement

2026-05-22 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

MonoScale is a novel framework designed to prevent performance collapse in LLM-based multi-agent systems (MAS) when new, heterogeneous agents are continuously integrated. It addresses the router's "cold start" problem by proactively generating a small set of agent-conditioned familiarization tasks. The system harvests evidence from both successful and failed interactions, distilling it into auditable natural-language memory to guide future routing decisions. Formalized as a contextual bandit, MonoScale employs trust-region memory updates, guaranteeing monotonic non-decreasing performance across onboarding rounds. Experiments on GAIA and Humanity's Last Exam (HLE) benchmarks demonstrate stable gains as the agent pool grows from 3 to 10 agents. A Qwen-3-30B-A3B-Instruct router with MonoScale achieved 55.15% accuracy on GAIA and 19.90% on HLE, outperforming naive scale-up and competitive with strong baselines like GPT-5 and Gemini-3-Pro, even in noisy agent environments.

Key takeaway

For AI Engineers building evolving multi-agent systems, integrating new agents demands a structured onboarding process. You should implement agent-conditioned warm-up tasks and memory-based router updates to prevent performance degradation. This approach ensures stable capability growth and robustness, even when dealing with unreliable agents or expanding agent pools, allowing your systems to scale reliably.

Key insights

Monotonic performance scaling in multi-agent systems requires proactive agent familiarization and memory-guided router updates.

Principles

• Naive MAS expansion causes performance collapse.
• Router cold-start requires execution-grounded knowledge.
• Trust-region memory updates ensure monotonic gains.

Method

Synthesize agent-conditioned warm-up tasks, collect success/failure traces, distill evidence into auditable natural-language memory, and apply trust-region policy optimization.

In practice

• Generate ~50 warm-up tasks per new agent.
• Distill positive and negative routing principles.
• Implement conservative fallback for new agents.

Topics

• Multi-Agent Systems
• LLM Routing
• Agent Onboarding
• Performance Scaling
• Contextual Bandits
• Trust-Region Optimization
• GAIA Benchmark

Best for: Research Scientist, AI Architect, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Building AI Agent Systems and Scaling Challenges in Agentic AI — Scaling AI agents requires architectural decomposition into multi-agent systems to manage complexity and prevent failure propagation.
• Scaling Behavior of Single LLM-Driven Multi-Agent Systems — LLM-driven multi-agent system performance shows diminishing returns due to coordination overhead, not monotonic scaling.
• ReM-MoA: Reasoning Memory Sustains Mixture-of-Agents Scaling — Structured cross-layer reasoning memory and curated diversified routing are critical for scalable Mixture-of-Agents performance.
• Scaling Managed Agents: Decoupling the brain from the hands — Decoupling AI agent components into stable interfaces enables adaptability to evolving models and scalable, resilient operations.
• The Hard Lessons from Running Dozens of AI Coding Agents in Parallel — Scaling AI coding agents amplifies single-agent failures, necessitating structured fixes and human strategic oversight.
• The Agentic Expansion Cascade — The agentic expansion is a substrate-driven consequence of parallel AI scaling laws, not a product strategy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.