Awesome Graph Engineering · 开放实践指南

工程化的不只是 单个智能体， 而是整个组织。

面向可编程 AI 智能体组织的精选实践指南——涵盖角色、拓扑、交接、工作图、验证、可靠性与可观测性。

浏览资源图谱 在 GitHub 上查看

在这里，图的节点是 AI 智能体，而不是数据实体。若你寻找的是图数据库、GNN、知识图谱或 GraphRAG， 请查看边界指南。

A graph of specialized AI agents connected by explicit handoffs and an evidence gate.

Focus the graph, then use arrow keys to orbit, plus or minus to zoom, and zero to reset. Hold Alt while scrolling to zoom without trapping normal page scrolling.

Working definition · emerging practice

30 秒理解 Graph Engineering

Graph Engineering 是对图结构智能体系统进行规范、执行、观测和演进的工程实践；它涵盖角色与运行时实例、连接它们的契约、共享状态与产物，以及评判集体行为的证据。合格系统至少运行两个可独立问责的智能体运行时实例；它们可以共享角色或模型。每个智能体可以包含本地循环；测试、审计循环、人类与现实世界锚点可以作为非智能体控制节点。该图必须实质性约束执行，并保持可检查。

  1. 01

    Node

    A bounded role specification or an accountable agentic runtime instance; only runtime instances satisfy plural agency.

  2. 02

    Edge

    A designed dependency: what crosses, in what shape, under which condition.

  3. 03

    Org graph

    The stable topology of roles, ownership, and standing relationships.

  4. 04

    Work graph

    The task-specific structure created, routed, and revised for one objective.

  5. 05

    Gate

    An independent test, audit, human decision, or real-world anchor that controls an edge.

该术语仍在形成且存在多种含义，因此这里将范围明确限定为智能体编排、多智能体系统和持久工作流中的实用工程层；它并非行业标准。

阅读完整定义与范围

The core model

Two graphs. One system.

A stable organization of agents executes ephemeral work. These are complementary analytical views—not proposed universal standards or necessarily separate runtime objects.

Drag or use arrows · Alt-scroll or +/− to zoom
Switch between a stable organization graph and an ephemeral work graph.

Use the Org graph and Work graph buttons to change the model. Focus the graph, then use arrow keys to orbit, plus or minus to zoom, and zero to reset.

Agent persistent role Edge contract explicit handoff Evidence gate quality or policy check

A practical taxonomy

Nine layers from roles to evolution

Start with roles, topology, and handoffs. Add state, gates, and operations only as real failure modes earn the complexity.

Before you add a second agent

Does this job need a graph?

Graphs add coordination cost, latency, and new failure modes. They earn that overhead only when at least one condition is real.

  1. 01

    Parallelism

    Can meaningful work proceed concurrently without constant coordination?

  2. 02

    Isolation

    Must contexts, permissions, or failure domains remain separated?

  3. 03

    Specialization

    Do distinct skills, tools, models, or durable contexts pay for separate roles?

  4. 04

    Verification

    Must a producer’s output pass an independent evidence gate before it advances?

None applyBuild a better loop.

Keep the free context-sharing and lower coordination cost of one agent.

One or more applyDesign the smallest graph that earns its overhead.

Start minimal, define the edges, and require evidence before you evolve.

Guided entry points

Start where you are

Choose the question closest to the one blocking you now.

探索交互式模型
01Loop builderI have one agent that works. Now what?

Decide whether a second role earns its place and draw the first edge.

 02ArchitectWhich topology and framework?

Match the shape of the graph to the shape of the task.

 03OperatorMake the graph survivable.

Design retries, evidence gates, budgets, tracing, and escalation.

 04ResearcherWhat does the evidence say?

Read the foundations, benchmarks, topology studies, and critiques.

Searchable resource directory

The graph engineering atlas

Search the field by design layer, resource type, and source/evidence label. Every result maps back to the README and open dataset.

Loading resources…

    Open data

    Use the field guide as data.

    Every resource ships as CSV and JSONL, mapped to a design layer and checked against the README.

    Download CSV Download JSONL View JSON Schema Explore on Hugging Face
    Python
    import pandas as pd

url = "https://raw.githubusercontent.com/ChaoYue0307/awesome-graph-engineering/main/data/resources.csv"
resources = pd.read_csv(url)

resources.query("layer == 'Gates'").head()
    Schemalayer stringrtype stringvenue stringyear integerdescription stringevidence string

    许可

    本仓库原创的元数据、数据结构、摘要、文档、代码和视觉资产均依据 CC0 1.0 Universal 贡献至公共领域。

    外部链接的论文、软件、名称、标志及其他第三方材料仍受其各自权利和许可证约束。CC0 不放弃商标权或专利权，且不对作品提供任何保证。欢迎为学术溯源进行引用，但 CC0 不要求引用。

    阅读完整许可