MonoScale: Scaling Multi-Agent System with Monotonic Improvement

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.

• The paper introduces MonoScale, a protocol that enforces monotonic improvement during sequential agent onboarding using memory-guided, trust-region updates.
• It employs agent-conditioned familiarization tasks and structured natural-language memory to mitigate cold-start misrouting and brittle workflow dependencies.
• Empirical results on the GAIA benchmark demonstrate that MonoScale prevents performance collapse and enhances system accuracy as the agent pool expands.

MonoScale: Expansion-Aware Sequential Scaling of LLM-Based Multi-Agent Systems

Motivation and Problem Formalization

The paper "MonoScale: Scaling Multi-Agent System with Monotonic Improvement" (2601.23219) targets a foundational challenge in the deployment of LLM-based multi-agent systems (MAS): how to expand the agent pool—by continuously integrating new specialized agents or tools—without risking performance collapse due to cold-start misrouting, heterogeneity, and unreliability in new agents. Classic orchestration assumes a static agent set, optimizing routing and coordination without addressing performance safety under continual agent augmentation. MonoScale offers a unified formalization for dynamic MAS expansion, casting sequential agent onboarding as a contextual bandit problem, and enforcing conservative policy updates via auditable natural-language memory.

Empirically, the authors demonstrate that naive expansion, simply introducing agents to the routing pool, can degrade end-to-end system accuracy on complex task workflows, as evidenced by results where increasing agents without adaptation leads to accuracy collapse (e.g., DeepSeek-V3.2 drops from 0.558 at 5 agents to 0.491 at 10 on GAIA). The core failure mode is cold-start router confusion: without grounded experience of the new agent's strengths, limitations, and failure modes, task dispatch amplifies unreliability and brittle workflows.

MonoScale Method: Expansion-Aware Familiarization and Memory Update

MonoScale proposes an expansion protocol with two key components:

1. Agent-Conditioned Familiarization Tasks: When onboarding a new agent, a custom set of warm-up probe tasks is synthesized to exercise the agent’s unique capabilities, interface boundaries, and typical error patterns, rather than relying on fixed suites or user traffic to surface failure modes. This agent-centric data synthesis leverages planner-executor-validator loops conditioned on agent cards, ensuring the router experiences a representative interaction subspace.
2. Structured Auditable Natural-Language Memory: Execution traces (including both successful and failed interactions) from the warm-up tasks are distilled into a structured memory for the router, encoding actionable routing principles, negative constraints, and exception clauses. This memory is editable and rollbackable, incorporated into the router’s prompt/context, and bounded by trust-region constraints that prevent abrupt behavioral shifts or erroneous policy updates.

Policy optimization is performed at the level of memory editing—rather than parameter tuning—using trust-region methods (KL-divergence constraints) to ensure monotonic non-decreasing performance across expansion rounds. Critically, each onboarding round includes a conservative fallback (disabling the new agent), ensuring regressions can be rolled back, and the deployed system’s reward remains safely bounded below by pre-expansion performance.

Theoretical Guarantee: Monotonic Non-Decreasing Performance

The orchestration process under sequential agent augmentation is modeled as a sequence of contextual bandit problems. The key theoretical result is a monotonicity theorem: under conservative lift and trust-region constrained memory updates, MonoScale provably ensures that the deployed router’s expected reward cannot decrease across expansion stages. For every expansion step $k$, with agent pool $S_k$ and memory $m_k$, the following guarantee holds:

$J_k(\pi_{m_k}) \geq J_{k-1}(\pi_{m_{k-1}})$

where $J_k$ denotes expected system reward under policy $\pi_{m_k}$ on agent pool $S_k$. The fallback constraint also provides a formal performance safety bridge, ensuring new agent integration does not compromise existing system functionality.

Empirical Results

Experiments are performed on the GAIA benchmark (tool use, planning, open-world reasoning) and the MCQ subset of Humanity’s Last Exam (deep reasoning), scaling agent pools from 3 to 10 agents:

• Strict Monotonicity: Under MonoScale, as agent pool size increases and routing memory is updated, overall accuracy on GAIA rises (e.g., Qwen-3-30B-A3B-Instruct improves from 44.84% for 3 agents to 55.15% for 10 agents), outperforming naive scale-up and matching or surpassing strong fixed-pool SOTA routers (GPT-5, Gemini-2.5-Pro).
• Robustness to Noisy Pools: In settings where agents exhibit malfunctioning or deceptive behaviors (semantic mismatch, failed tools, false capability advertising), MonoScale maintains or slightly improves overall performance as the pool grows, in stark contrast to performance collapse on strong router baselines without memory updates (Gemini-3-Pro shows a drop from 65.45% to 52.73% as agents increase from 5 to 7).
• Practical Scalability: MonoScale is shown to be effective for agent pools up to 10, and memory-guided routing enables smaller open-weight models to rival proprietary giants on complex reasoning and tool workflows.

Analysis of Failure Modes and Resolution

Case studies in the appendix elucidate two archetypal failure modes addressed by MonoScale:

• Cold-Start Misjudgment: Without execution-grounded experience, routers may erroneously delegate tasks to new agents based on superficial description matching, amplifying failures. MonoScale’s familiarization phase exposes these boundaries before deployment.
• Brittle Workflow Links: Specialized agents can introduce rigid dependencies or lose essential context, causing single-point workflow failure. MonoScale’s memory distills interface boundaries and coordination patterns, guiding safe agent selection and context passing.

Resolution is achieved by routing backed by memory constraints, which encode both positive and negative evidence extracted during warm-up. This enables the router to actively avoid misassignment, leverage multimodal tools at the proper step, and suppress unreliable dependencies.

Implications and Future Directions

MonoScale’s agent onboarding protocol establishes a scalable blueprint for MAS evolution in dynamic, open-world settings such as the Agentic Web, where systems must continuously integrate third-party agents. The protocol achieves verifiable safety, robustness to unreliable or deceptive agents, and practical gains for open-source models.

Several important directions emerge:

• Scalability to Million-Agent Catalogs: Scaling MonoScale beyond tens to thousands or millions of agents (where routing is retrieval-driven) will necessitate budgeted retrieval-calibration loops and parallelized onboarding.
• Integration with Adversarial Onboarding: Security, privacy, and robustness considerations (managing prompt-injection, evidence poisoning, sensitive data retention) must augment the protocol, including adversarial test suites and sandboxed permissions.
• Dynamic Memory and Retrieval-Driven Routing: Efficient memory management (pattern retrieval, compact storage), budget-aware agent selection, and continuous calibration for long-tail agents are open problems.

Conclusion

MonoScale provides a principled, theoretically grounded, and empirically validated protocol for scaling LLM-based multi-agent systems with guaranteed monotonic performance under sequential agent augmentation. By combining agent-conditioned warm-up tasks and memory-based conservative updates, the framework mitigates cold-start and performance collapse risks associated with naive expansion, and lays a foundation for robust, scalable agent orchestration in dynamic open-world environments.