When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail

TL;DR

This work investigates whether a single-agent system endowed with a library of skills (SAS) can match the performance of traditional multi-agent systems (MAS) while reducing computational cost. It formalizes MAS-to-SAS compilation as a three-phase process that decomposes agent capabilities, assigns backends, and internalizes topology, enabling a faithful, cost-efficient transformation with substantial token and latency savings. A central contribution is the Skill Scaling Hypothesis, which predicts a non-linear phase transition in skill selection accuracy governed by a capacity threshold κ and driven by semantic confusability among skills, rather than library size alone; hierarchical routing is shown to mitigate scaling degradation. The findings connect cognitive science principles—Hick’s Law, cognitive load theory, similarity-based interference, and chunking—to LLM-based action selection, and provide practical guidelines for designing scalable skill libraries, including hierarchy-based organization and careful descriptor engineering. Overall, the work offers a cognitive-grounded framework and empirical evidence for the limits and opportunities of using skill-based single-agent agents as scalable alternatives to multi-agent coordination in large language model systems.

Abstract

Multi-agent AI systems have proven effective for complex reasoning. These systems are compounded by specialized agents, which collaborate through explicit communication, but incur substantial computational overhead. A natural question arises: can we achieve similar modularity benefits with a single agent that selects from a library of skills? We explore this question by viewing skills as internalized agent behaviors. From this perspective, a multi-agent system can be compiled into an equivalent single-agent system, trading inter-agent communication for skill selection. Our preliminary experiments suggest this approach can substantially reduce token usage and latency while maintaining competitive accuracy on reasoning benchmarks. However, this efficiency raises a deeper question that has received little attention: how does skill selection scale as libraries grow? Drawing on principles from cognitive science, we propose that LLM skill selection exhibits bounded capacity analogous to human decision-making. We investigate the scaling behavior of skill selection and observe a striking pattern. Rather than degrading gradually, selection accuracy remains stable up to a critical library size, then drops sharply, indicating a phase transition reminiscent of capacity limits in human cognition. Furthermore, we find evidence that semantic confusability among similar skills, rather than library size alone, plays a central role in this degradation. This perspective suggests that hierarchical organization, which has long helped humans manage complex choices, may similarly benefit AI systems. Our initial results with hierarchical routing support this hypothesis. This work opens new questions about the fundamental limits of semantic-based skill selection in LLMs and offers a cognitive-grounded framework and practical guidelines for designing scalable skill-based agents.

Figures (8)

• Figure 1: Skill-based agents: efficiency gains and scaling limits. (a) Compiling multi-agent systems into single-agent skill libraries reduces communication overhead, cutting latency and token usage. (b) Skill selection accuracy degrades non-linearly as libraries grow, exhibiting a phase transition at a capacity threshold. As skill libraries grow, the increased size and semantic confusability among skills drive this degradation; hierarchical routing restores reliable selection by organizing skills into structured categories. Visualization modified from Nana Banana.
• Figure 2: Scaling law fit quality. The proposed functional form $\textsc{Acc} \approx \alpha/(1+(|\mathbf{S}|/\kappa)^\gamma)$ achieves excellent fit ($R^2 > 0.97$) for both models, validating the theoretical model.
• Figure 3: Effect of skill competitors on selection accuracy. Green: no competitors (each skill unique). Orange: 1 competitor per skill. Red: 2 competitors per skill. At fixed total library size, higher confusability leads to lower accuracy, demonstrating that semantic similarity—not library size alone—drives selection errors.
• Figure 4: Effect of execution policy complexity on selection accuracy. Each panel shows results for one model. Contrary to expectations, the three complexity levels show largely overlapping performance curves.
• Figure 5: Effect of hierarchical routing on selection accuracy. Blue: flat selection. Red: naive domain hierarchy. Green: confusability-aware hierarchy. At large library sizes ($|\mathbf{S}| \geq 60$), hierarchy maintains $\sim$72--85% accuracy while flat selection degrades to $\sim$45--63%.

Theorems & Definitions (12)

• Definition 2.1: Agent
• Definition 2.2: Multi-Agent System
• Definition 2.3: Skill
• Remark 2.1: Separation of Selection and Execution
• Definition 2.4: Single-Agent with Skills
• Remark 2.2
• Definition 2.5: Compilation
• Definition 2.6: Behavioral Fidelity
• Definition 2.7: Cost Efficiency
• Definition 3.1: Compilability