Researchers simplify increasingly complex AI problem-solving, one step at a time

A team of researchers at Tsinghua University has proposed a mathematical framework for building and understanding increasingly complex artificial intelligence systems by breaking difficult tasks into smaller, coordinated components. The approach, called the “calculus of intelligence,” or COIN, is designed for agentic AI—systems capable of planning and acting with limited human supervision in areas such as cybersecurity, software development and automated decision-making.

The framework is based on a simple but powerful idea: complex intelligence may be easier to create when it is assembled from many specialized forms of limited intelligence. Rather than asking one enormous AI model to understand and solve an entire problem at once, COIN divides the overall objective into subtasks that can be analyzed, executed and verified separately. The results are then recombined into a coherent solution that satisfies the requirements of the original task.

Yang Yuan, an associate professor at the Institute for Interdisciplinary Information Sciences at Tsinghua University and corresponding author of the study, compared the concept with classical calculus. Calculus can determine the area beneath a complicated curve by dividing it into many smaller sections and adding their contributions together. In a similar way, COIN seeks to represent a complex intelligent process as a structured composition of simpler operations.

The mathematical foundation of the framework is a structure known as a Grothendieck topos. In mathematics, a topos can be understood as a formal environment that describes objects, relationships and the rules governing how they interact. For artificial intelligence, the researchers use this setting to specify what each component of a system is allowed to observe, which conditions it must satisfy and how it should exchange information with other components.

Yuan likened the arrangement to the design of a large aircraft. Engineers responsible for wings, engines, navigation and control systems work on different parts of the airplane and do not need access to every detail of the entire project. Each group operates within a limited local view, but the components must still fit together at shared boundaries. The wing must connect correctly to the fuselage, the engines must interact with the control system and all parts must meet common safety requirements.

In COIN, these local views correspond to specialized subtasks or agents. The framework establishes the interfaces through which they communicate and defines how their solutions can be combined without violating global constraints. This is intended to address one of the central challenges facing agentic AI: coordinating independent systems while preserving consistency, reliability and accountability across the complete workflow.

The researchers describe the process using the language of decomposition and recomposition. Decomposition breaks a broad objective into smaller problems that are sufficiently well-defined for individual models or agents to solve. Recomposition then assembles those local results into a global outcome. The mathematical rules are important because simply joining the outputs of multiple AI systems can produce contradictions, duplicated work or failures at the points where their responsibilities overlap.

According to Yuan, the framework reflects a broader view of intelligence in which intelligence is inseparable from structure. A system may appear incomprehensible because its organization is hidden, rather than because its individual parts are intrinsically impossible to understand. If the correct structure can be identified, a seemingly overwhelming problem may become a collection of bounded tasks that can be tested independently. This could make future AI systems easier to inspect and verify, particularly when they are deployed in high-stakes environments.

The proposal also challenges the idea that progress in AI must depend on creating a single model with unlimited capabilities. Many smaller systems, each possessing a narrow but useful form of intelligence, could potentially be organized into networks capable of solving problems beyond the capacity of any one model or human expert. Such systems might distribute planning, perception, reasoning, coding and monitoring across specialized agents while using formal interfaces to maintain cooperation.

COIN is presented as an early step toward a common mathematical language for intelligence. The researchers aim for such a language to describe what AI models learn, how complex tasks can be divided and how numerous limited intelligences can be coordinated into larger systems. The work, co-authored by Andrew Chi-Chih Yao, professor and dean of Tsinghua’s Institute for Interdisciplinary Information Sciences, appears in the journal iFuture under the title “Calculus of intelligence: A topos-monadic framework for agentic workflows.” If the framework can be translated into practical engineering tools, it could influence how future AI systems are designed—not as solitary superintelligences, but as carefully coordinated communities of specialized agents.

News Publication Date:
17-Jul-2026

Web References:
https://doi.org/10.26599/IF.2026.9710001
https://www.sciopen.com/journal/3135-3169

References:
Yang Yuan and Andrew Chi-Chih Yao, “Calculus of intelligence: A topos-monadic framework for agentic workflows,” iFuture, DOI: 10.26599/IF.2026.9710001.

Keywords

Artificial intelligence, agentic AI, calculus of intelligence, COIN, Grothendieck topos, mathematical AI, AI agents, multi-agent systems, task decomposition, intelligent systems, Tsinghua University, AI workflows