AI Skills AI技能 3h ago Updated 2h ago 更新于 2小时前 50

Graph Engineering: Engineering Coordination Instead of Smarter Agents 图工程:工程化协调而非更智能的代理

Reliable agent systems depend on contracts governing inter-agent interactions rather than the number of agents. Graph engineering focuses on coordinating multiple units of work through explicit control over state, routing, ownership, evidence, and authority. Not every node in a system should be an agent; deterministic software is often better for tasks like routing, validation, and permissions. Typed state schemas separate different types of information (facts, proposals, decisions, artifacts) t 核心发现:可靠的多智能体系统不取决于智能体的数量,而取决于定义它们之间交互的契约(Contracts)。 技术贡献:提出了“图工程”(Graph Engineering)的概念,将其与“循环工程”区分开来,强调通过显式的状态、路由和所有权控制来协调多个工作单元。 重要性:解决了多智能体系统中的“协作鸿沟”,指出可靠性更多依赖于节点间的交接形状而非单个模型的能力。 实践建议:不应将所有节点都设计为智能体,确定性软件更适合处理路由、验证等任务;应使用类型化状态模式分离事实、提案和决策。 架构视角:图中的每一条边都应被视为一个包含证据、权限、路由和预算的契约,而不仅仅是流程图的箭头。

70
Hot 热度
75
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • Reliable agent systems depend on contracts governing inter-agent interactions rather than the number of agents.
  • Graph engineering focuses on coordinating multiple units of work through explicit control over state, routing, ownership, evidence, and authority.
  • Not every node in a system should be an agent; deterministic software is often better for tasks like routing, validation, and permissions.
  • Typed state schemas separate different types of information (facts, proposals, decisions, artifacts) to ensure trust and clarity across nodes.
  • Edges in a graph represent operational contracts that define what structured state moves, what evidence is required, and which budget or permissions are inherited.

Why It Matters

This article highlights the critical importance of designing robust coordination mechanisms in multi-agent systems, which is essential for building reliable and scalable AI applications. By emphasizing the need for clear contracts and typed state schemas, it provides a framework for engineers to avoid common pitfalls such as ambiguous handoffs and untrustworthy state propagation. Understanding these principles can significantly improve the reliability and maintainability of complex agent-based systems.

Technical Details

  • Graph Engineering: This concept involves designing the topology of nodes, transitions, state, and dependencies in a multi-agent system. It ensures that each interaction between agents is governed by clear contracts that specify what data is passed, who owns the next state, and what permissions are inherited.
  • Node Types: The article distinguishes between different types of nodes, including probabilistic actors (agents), deterministic functions, evaluators, and human gates. Each node should have a distinct objective, permission set, input/output contract, or independent verifier.
  • Typed State Schema: A recommended approach to managing state in multi-agent systems is to use a typed schema that separates facts, proposals, decisions, artifacts, and budgets. This ensures that different types of information are handled with appropriate levels of trust and clarity.
  • Edge Contracts: Each edge in the graph represents a detailed contract that specifies the structured state moving between nodes, the evidence required, the route type, and the budget or permissions inherited by the destination. This helps in maintaining clear boundaries and reducing ambiguity.

Industry Insight

  • Shift from Agent-Centric to Contract-Centric Design: The industry should move towards a design philosophy that emphasizes clear contracts and well-defined interfaces between agents rather than simply adding more agents. This shift will lead to more reliable and maintainable systems.
  • Integration of Deterministic Software: There is a growing recognition that deterministic software components are often more effective for tasks like routing, validation, and permissions management compared to probabilistic agents. Integrating these components can enhance the overall performance and reliability of AI systems.
  • Standardization of State Management: The adoption of standardized state management practices, such as using typed state schemas, can help in reducing complexity and improving interoperability between different parts of a multi-agent system. This standardization is crucial for scaling AI applications effectively.

TL;DR

  • 核心发现:可靠的多智能体系统不取决于智能体的数量,而取决于定义它们之间交互的契约(Contracts)。
  • 技术贡献:提出了“图工程”(Graph Engineering)的概念,将其与“循环工程”区分开来,强调通过显式的状态、路由和所有权控制来协调多个工作单元。
  • 重要性:解决了多智能体系统中的“协作鸿沟”,指出可靠性更多依赖于节点间的交接形状而非单个模型的能力。
  • 实践建议:不应将所有节点都设计为智能体,确定性软件更适合处理路由、验证等任务;应使用类型化状态模式分离事实、提案和决策。
  • 架构视角:图中的每一条边都应被视为一个包含证据、权限、路由和预算的契约,而不仅仅是流程图的箭头。

为什么值得看

这篇文章对AI从业者极具意义,因为它指出了当前多智能体系统落地时最容易被忽视的痛点:随着系统复杂度增加,瓶颈从模型能力转移到了系统架构的协作机制上。它提供了一种新的工程化视角(图工程),帮助开发者构建比单纯堆砌智能体更可靠、可维护的生产级系统。

技术解析

  • 循环与图的区别:文章定义了“Loop Engineering”负责单一工作单元的收敛(如重试修复),而“Graph Engineering”负责多个工作单元之间的协调。生产环境中的图通常包含多个局部循环,但图本身决定了这些循环的边界、允许变更的内容及最终完成权的归属。
  • 五要素词汇体系:为了厘清概念,文章将常被混淆的术语拆解为五个独立实体:Agent(概率性执行者)、Node(任何工作单元,包括模型调用或人类审核)、Loop(有界反馈周期)、Graph(节点、转换、状态和依赖的拓扑结构)、Governor(在节点分歧或高风险时拥有决策权)。
  • 非智能体节点的必要性:明确指出并非所有节点都应是智能体。路由、Schema验证、权限检查和预算执行等功能由确定性软件实现更为高效、便宜且不易出错,因为智能体可能会在模糊的交接中自信地误读指令。
  • 边缘即契约(Edge as Contract):工程化的图中,每条连接线(Edge)不仅是流程顺序,更是一个操作契约。它明确规定了移动的结构化状态、所需证据、下一状态的归属权、路由类型以及继承的预算和权限。例如,研究员到实施员的交接需明确输入数据、触发条件、修改权限和时长限制。
  • 状态管理的类型化方案:针对多节点协作中最难的状态一致性问题,文章反对共享全局对话历史,主张使用类型化状态模式(Typed State Schema)严格分离 Facts(事实)、Proposals(提案)、Decisions(决定)、Artifacts(产物)和 Budgets(预算)。文章提供了Python TypedDict示例,展示了如何通过代码结构强制区分不同信任级别的数据,防止推测性笔记被误认为既定事实。

行业启示

  • 架构重心转移:AI系统工程的重心将从“寻找更聪明的模型”转向“设计更健壮的编排架构”。未来的竞争壁垒在于如何定义和管理智能体间的接口契约与状态流转,而非单纯的参数量或推理速度。
  • 混合架构成为标准:成熟的AI系统将采用“智能体 + 确定性逻辑”的混合架构。对于需要高一致性和低延迟的控制流(如审批、校验、路由),应优先使用传统软件工程方法,仅在需要创造性或不确定性处理的环节引入智能体。
  • 标准化契约语言的缺失与机遇:目前缺乏统一的图工程标准语言来描述这些交接契约。这预示着未来可能出现专门用于定义多智能体系统状态机、权限和数据传输格式的新兴工具层或中间件框架,以降低分布式AI系统的开发门槛。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent LLM 大模型 Programming 编程