A group of researchers spanning universities in China, Hong Kong and the United States has published a resource library, hosted on GitHub as Awesome-Graph-Engineering, that catalogs papers, benchmarks and open-source tools for building multi-agent AI systems around graph structures rather than linear prompt chains. The collection accompanies a survey titled “Graph Engineering in the Era of LLM Agents,” posted to arXiv in August. It is maintained by a team affiliated with DEEP-JLU, a lab at Jilin University.
The survey’s central argument is that AI system design has moved through three distinct layers of engineering effort, and each layer demands a different structure. Model Intelligence covers the work of building and adapting a foundation model itself: pretraining, post-training, prompt engineering and context engineering. Individual Intelligence covers what a single agent needs to act on its own: tool use, memory, skill composition, and a runtime loop. System Intelligence, the newest and least standardized layer, covers what happens once many agents have to share tasks, state, and history at the same time. The authors argue that only an explicit graph, not a sequence of prompts or a fixed pipeline, can represent that kind of shared, evolving structure.
That framing matters because most production agent stacks were not built with a graph as the native data structure. LangChain, the OpenAI Agents SDK, and the Claude Agent SDK, all cataloged here under Individual Intelligence, are designed around a single agent’s loop. The libraries the survey places under System Intelligence instead model a multi-agent workflow as nodes and edges: who talks to whom, what state persists between turns, and how the topology itself can change mid-task. That group includes LangGraph, Microsoft’s Agent Framework, Google’s ADK, AutoGen’s GraphFlow, CrewAI, and CAMEL. Graphiti, listed separately under memory, applies the same graph logic to what an agent remembers across sessions rather than to how agents coordinate.
The collection also traces where these ideas already ship in production-shaped systems rather than staying confined to benchmarks. Coding agents such as MetaGPT, SWE-agent, OpenHands, and Claude Code sit under a software-engineering application category. Scientific-discovery tools like SciAgents and The AI Scientist, clinical systems such as AMIE, and several enterprise workflow platforms round out the applications list, each treated as a case study in how graph-structured coordination behaves once it leaves the lab.
The repository does not include independent benchmark comparisons between the graph-native frameworks it lists; it presents them as a taxonomy, not a leaderboard.
For a builder deciding whether to adopt a graph-based orchestration framework instead of chaining single-agent calls, this repository reads closer to a literature review than a tutorial. It will not say which library ships fastest. It does say which failure modes, among them state loss across agent handoffs, brittle task decomposition, and the absence of a shared ontology between agents, the field has already spent effort naming, along with which open-source projects claim to address each one. Teams currently stitching multi-agent workflows together with ad hoc queues and shared scratch files should treat the System Intelligence section as a checklist of what a graph-native runtime would need to replace before a rebuild is worth the cost.
Collection and survey published by the DEEP-JLU research group on GitHub as Awesome-Graph-Engineering, accompanying an arXiv preprint posted in August 2026.
