展示HN:IntentusNet - 一个用于多智能体工作流的安全意图路由器和运行时环境

1作者: balachandarmani2 个月前原帖
嗨,HN, 在过去的几个月里,我一直在开发IntentusNet,这是一个小型、语言无关的运行时,用于在代理和工具之间路由“意图”,并支持可选的加密和多种传输方式。 GitHub: [https://github.com/Balchandar/intentusnet](https://github.com/Balchandar/intentusnet) ### 这个项目试图解决的问题 我看到的大多数多代理/工具调用设置最终都变成了大量的临时拼凑: - 各个代理之间的自定义消息格式 - 手动编写的路由和回退逻辑 - 在一个地方使用HTTP,在另一个地方使用WebSocket,可能在其他地方使用ZeroMQ - 没有一致的追踪或错误模型 - 没有明确的地方添加安全性(加密、来源、身份链) MCP非常适合描述工具,但它并不试图成为一个运行时或路由器。我想要一些可以在MCP及其他技术栈之下或旁边工作的东西,能够简单回答: “给定一个意图,哪个代理应该处理它,使用哪种传输方式,采用什么回退机制,以及我们如何安全地封装它?” ### IntentusNet提供的功能(当前) 在核心部分,它有几个小的基本组件: - **IntentEnvelope** – 结构化消息,包含上下文、元数据、路由选项和标签 - **AgentRegistry** – 内存中代理和能力的注册表(它们处理哪些意图,可选的回退链) - **IntentRouter** – 选择一个代理,执行它,并在主要代理失败时应用回退 - **Transports** – 可插拔的传输方式,目前包括: - 进程内(直接路由调用) - HTTP(带有TransportEnvelope的POST请求) - WebSocket(双工、异步) - ZeroMQ(REQ/REP客户端加上简单服务器) - **Tracing** – 简单的追踪接收器,记录跨度(代理、意图、延迟、状态、错误) - **EMCL(加密模型上下文层)** – 可选的信封: - 一个基于HMAC的简单演示提供者 - 一个AES-GCM提供者(AES-256-GCM,base64,身份链)用于真正的加密 此外,还有一个MCP适配器,它接受MCP工具请求({ name, arguments }),将其封装到IntentEnvelope中,通过IntentusNet路由,并返回MCP风格的结果。这个想法是MCP工具可以通过同一个路由器连接,无论是否使用EMCL。 ### 示例用法 一个最小的流程如下: - 定义具有“summarize.document.v1”、“translate.text.v1”、“store.note.v1”等能力的代理 - 在运行时注册它们 - 使用IntentusClient发送一个带有有效载荷的意图;路由器选择合适的代理,并在配置的情况下处理回退 还有一个小型“助手”风格的设置示例,其中一个类似NLU的代理解析自然语言请求,并向专门的代理(如日历、地图等)发出下游意图,以展示多代理路由的实际应用。 ### 状态/缺失的内容 这是一个早期阶段的项目: - 运行时使用Python;其他SDK(例如C#)尚未准备好 - 编排/工作流层(序列、并行步骤、分支)在RFC中进行了初步勾勒,但仅部分实现 - 文档和示例确实可以改进 - 目前尚未声明生产就绪:这更像是一个结构化的参考实现或起始点 我非常希望能收到关于以下方面的反馈: - 架构(协议、路由器、传输、EMCL之间的分离是否合理?) - 这种类型的意图路由器在您正在构建的MCP或基于工具的系统中是否真的有用 - 您期望在这样的运行时中看到的缺失基本组件(超时、背压、更丰富的追踪等) - 在现实世界的工作流中,小型、传输无关、支持EMCL的路由器将如何提供帮助(或在什么情况下显得过于复杂) 如果您正在使用MCP、多代理设置或安全工具调用,并对这个项目的运作方式有意见,我非常希望听到您的想法——无论是在这里,还是通过GitHub上的问题或PR。
查看原文
Hi HN,<p>Over the last few months I’ve been working on IntentusNet, a small, language-agnostic runtime for routing “intents” between agents and tools, with optional encryption and multiple transports.<p>GitHub: <a href="https:&#x2F;&#x2F;github.com&#x2F;Balchandar&#x2F;intentusnet" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;Balchandar&#x2F;intentusnet</a><p>What problem this tries to solve<p>Most multi-agent &#x2F; tool-calling setups I’ve seen end up as a lot of ad-hoc glue:<p>- custom message formats between each agent - hand-rolled routing and fallback logic - HTTP in one place, WebSocket in another, maybe ZeroMQ somewhere else - no consistent tracing or error model - no clear place to add security (encryption, provenance, identity chain)<p>MCP is great for describing tools, but it doesn’t try to be a runtime or router. I wanted something that sits underneath or alongside MCP and other stacks, and just answers:<p>“Given an intent, which agent should handle it, through which transport, with what fallback, and how do we wrap it securely?”<p>What IntentusNet provides (today)<p>At the core it has a few small primitives:<p>- IntentEnvelope – structured message with context, metadata, routing options, and tags - AgentRegistry – in-memory registry of agents and capabilities (which intents they handle, optional fallback chains) - IntentRouter – picks an agent, executes it, and applies fallback if the primary fails - Transports – pluggable transports, currently: - in-process (direct router call) - HTTP (POST with a TransportEnvelope) - WebSocket (duplex, async) - ZeroMQ (REQ&#x2F;REP client plus simple server) - Tracing – simple trace sink that records spans (agent, intent, latency, status, error) - EMCL (Encrypted Model Context Layer) – optional envelope: - a simple HMAC-based demo provider - an AES-GCM provider (AES-256-GCM, base64, identity chain) for real encryption<p>There’s also an MCP adapter that takes an MCP tool request ({ name, arguments }), wraps it into an IntentEnvelope, routes it through IntentusNet, and returns an MCP-style result. The idea is that MCP tools can be wired through the same router, with or without EMCL.<p>Example usage<p>A minimal flow looks like:<p>- define agents with capabilities like “summarize.document.v1”, “translate.text.v1”, “store.note.v1” - register them in the runtime - use the IntentusClient to send an intent with a payload; the router picks the right agent and handles fallback if configured<p>There’s also an example of a small “assistant”-style setup where an NLU-like agent parses a natural language request and emits downstream intents to specialized agents (calendar, maps, etc.), just to show multi-agent routing in practice.<p>Status &#x2F; what’s missing<p>This is early-stage:<p>- Runtime is in Python; other SDKs (for example C#) are not ready yet - Orchestrator &#x2F; workflow layer (sequences, parallel steps, branching) is sketched out in RFCs but only partially implemented - Docs and examples can definitely be improved - No claims of production readiness yet: it’s more of a structured reference implementation or starting point<p>I’d really appreciate feedback on:<p>- the architecture (does the separation between protocol, router, transports, EMCL make sense?) - whether this kind of intent router is actually useful under MCP or tool-based systems you’re building - missing primitives you’d expect in a runtime like this (timeouts, backpressure, richer tracing, etc.) - real-world workflows where a small, transport-agnostic, EMCL-capable router would help (or where it’s overkill)<p>If you’re working with MCP, multi-agent setups, or secure tool calling and have opinions on how this should work, I’d love to hear them—either here or via issues or PRs on GitHub.