Large language models (LLMs) have evolved from simple language generators into autonomous agents capable of tackling complex, long-horizon tasks. This evolution has spawned several engineering paradigms, including Prompt Engineering, Context Engineering, Harness Engineering, and Loop Engineering, each addressing different aspects of agent capability. However, as tasks grow in complexity, individual intelligence hits a fundamental limit: many tasks require heterogeneous expertise, interdependent subtasks, parallel execution, independent verification, and persistent state—all of which exceed the organizational capacity of any single agent.
To address this, researchers from Hugging Face and Jilin University introduce Graph Engineering, a new paradigm for building next-generation agent systems. Unlike previous approaches that optimize individual interactions or agent-level behaviors, Graph Engineering constructs explicit, dynamic, and evolving graph structures that represent tasks, agents, and system states. These graph-based abstractions provide a unified foundation for organizing complex objectives, orchestrating heterogeneous agents, modeling system dynamics, and enabling scalable agent evolution.
The paper argues that achieving System Intelligence—the ability of an agent system to coordinate multiple intelligent components into a coherent, adaptive whole—requires more than just adding more agents. It demands explicit structures for organizing work, coordinating diverse agents, and maintaining evolving execution states. Graph Engineering offers exactly that, serving as a bridge from individual intelligence to system-level intelligence.
The survey systematically reviews the principles, methodologies, and applications of Graph Engineering for LLM agents. It also provides a comprehensive collection of related resources, including papers, open-source data, and projects, available at Awesome-Graph-Engineering.