Universal Primitives for Agentic Reasoning: A Formal Decomposition of LLM Agent Frameworks

The proliferation of large language model (LLM) agent frameworks—exceeding 200 distinct architectures by early 2026—has created a fragmentation crisis in agentic AI research. This paper presents a formal decomposition of LLM agent architectures into 13 universal cognitive primitives (Σ-primitives), derived through systematic analysis of over 60 human reasoning frameworks spanning education, evolutionary biology, decision science, and strategic management. Every major agentic AI framework examined—from ReAct and Chain-of-Thought to AutoGPT, LangChain, and Claude Code—can be expressed as ordered compositions of these primitives. The analysis reveals six recurring composition stacks, identifies systematic gaps in current agent designs, and provides structural evidence of convergent design between human cognitive architectures and artificial agent architectures.