One agent, one skill library: 54% cheaper than multi-agent — until it tops ~90 skills

Notes on When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail (arXiv:2601.04748) — Xiaoxiao Li · January 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill retrieval

Multi-agent systems split a task across several specialized agents that talk to each other, and that coordination is expensive — every hand-off costs tokens and latency. Compile the same pipeline into a single agent backed by a library of skills, with each former agent's behavior turned into one skill the agent selects from, and the inter-agent messages disappear. That's the substitution tested here, across GSM8K, HotpotQA, and HumanEval, then pushed two ways: the skill library grows from a handful of entries to dozens, and separately gets seeded with near-duplicate competitor skills to see what actually breaks selection.

Compilation works about as well as advertised: token use drops 53.7% on average, latency drops 49.5%, and accuracy lands within -2.0% to +4.0% of the original multi-agent system (average +0.7%). Skill selection, though, doesn't scale as cleanly. Accuracy holds above roughly 90% while the library stays small, then falls off sharply past a fitted capacity point — about 92 skills for GPT-4o-mini, 84 for GPT-4o — a phase transition rather than a gradual slide, echoing capacity limits reported in human decision-making research. Count isn't the driver; semantic confusability is. With zero near-duplicate skills, accuracy holds at 100% even at 20 skills; add one confusable competitor per skill and it drops 7-30%; add two and it drops 17-63%. A two-stage hierarchical routing scheme — group skills, then pick within the group — claws back most of that: +37 to +40 points for GPT-4o-mini at large library sizes, +9 to +10 points for GPT-4o.

Key numbers

Token usage vs. multi-agent baseline	-53.7% avg
Latency vs. multi-agent baseline	-49.5% avg
Capacity threshold (κ)	~84-92 skills
Accuracy drop, 1 confusable skill per entry	-7 to -30%
Hierarchical routing recovery (small model)	+37 to +40 pts

What this means for skill authors

The practical read for skill authors: how many skills sit in a library matters less than how distinguishable their descriptions are. A larger library of tightly-scoped, differently-worded skills can outperform a smaller one where entries overlap in phrasing — the failure mode here is confusion between similar options, not overload from sheer count. Generic, template-shaped descriptions, the kind that read like every other entry in a category, are exactly what gets misrouted.

For anyone maintaining a skill catalog, that argues for active curation over passive accumulation: deduplicating near-identical skills and forcing distinct, specific descriptions does more for selection accuracy than pruning by raw count ever would. It also backs splitting a catalog into categories rather than one flat list once it grows large — the hierarchical split recovered most of the accuracy lost to scale here. SkillFed's dual-tool scanning and vetted catalog exist for security and quality reasons first, but this hands curation a second, independent payoff as a library grows: fewer near-duplicate skills competing for the same selection.

Read the specific threshold loosely. The libraries here are synthetic, the study measures selection accuracy in isolation rather than whether the downstream task gets done right, and only two OpenAI models were tested — the number ~85-90 won't transfer as-is to other models or real-world catalogs. What should transfer is the shape of the failure: stable, then a cliff, driven by overlap between skills rather than their count.

Skills related to this research

design-flow design-flow walks you through the complete design process as a guided sequence, running each phase in order and waiting for confirmation before advancing. It orchestrates six core phases—from clarifying your thinking through building components—plus an optional design review, with the ability to skip phases you don't need and resume mid-project where you left off.★ 461 Nextjs React Implementation Transform design specification documents into fully-featured Next.js/React components complete with TypeScript types, Tailwind styling, browser-mode tests, and Storybook stories. This skill automates the journey from design to implementation, handling component structure, variant logic, responsive behavior, and accessibility patterns across atoms, molecules, and organisms.★ 3

Related notes

References

Xiaoxiao Li, "When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail," arXiv:2601.04748 (January 2026)
Hick, W. E., "On the Rate of Gain of Information" (1952) — Hick's Law, cited as the choice-reaction-time basis for the capacity limit
Sweller, J., Cognitive Load Theory (1988), cited as the basis for bounded selection capacity
Shepard, R. N., "Toward a Universal Law of Generalization for Psychological Science" (1987), cited as the basis for similarity-driven interference