DEEP-JLU/Awesome-Graph-Engineering

What it solves

As LLM agents move from single-model tasks to complex, coordinated systems, organizing their behavior becomes difficult. This project addresses that by collecting resources on graph engineering — using explicit, dynamic graph structures to plan tasks, coordinate multiple agents, track runtime state, and support system evolution. It gives researchers and builders a single starting point to understand and apply graph-based approaches across the full spectrum of agent development.

How it works

The repository is a curated, survey-aligned list. It organizes resources into three intelligence levels, mirroring its accompanying paper:

• Model Intelligence — resources on training, prompting, and context engineering (like chain-of-thought, RAG, and prompt optimization).
• Individual Intelligence — resources for making a single agent capable: tool integration, memory management, skill composition, runtime orchestration, and interaction loops.
• System Intelligence — resources for coordinating many agents with explicit graphs, including task organization, agent coordination, state management, system evolution, and ontology engineering.

Within each level, the README lists research papers, benchmarks, datasets, environments, open-source libraries, and real-world applications. The whole collection is meant to grow: it invites pull requests and issues for missing or new work.

Who it’s for

This is for researchers, engineers, and students working on LLM-based agents, multi-agent systems, graph-based reasoning, or agent infrastructure. If you are designing an agent workflow, building coordination layers, or exploring how graphs can structure agent state and evolution, this list gives you a broad, organized map of the field.

Highlights

• Taxonomy spans from model-level (pre-training, post-training, prompting) to system-level (coordination, state, evolution).
• Includes both academic papers and practical open-source libraries and benchmarks.
• Continuously updated; contributions are explicitly welcomed.
• Directly tied to an arXiv survey paper for deeper reading.