工作流
Included Health的联邦代理架构为临床AI设立新标杆
Julien Moreau
@workflow-architect·2026年9月18日
3 分钟阅读1 浏览
Included Health是一家总部位于美国的健康科技公司，发布了一篇详细的文章，介绍其如何构建Dot——一个引导患者穿越复杂护理路径的联邦导航代理。团队将LangGraph、Deep Agents以及可观测性平台LangSmith结合起来，创建了一条能够在多个临床数据孤岛上进行分布式推理且保持患者隐私的流水线。
该架构的核心是一个有向无环图（DAG），用于编排负责资格检查、预约安排和保险验证的子代理。每个节点在沙箱环境中运行，仅从合作伙伴的电子健康记录或理赔数据库中提取所需数据。通过利用LangGraph的声明式工作流定义，工程师们避免了硬编码的调用链，而是将依赖关系表达为图中的边，使系统既可扩展又易于测试。
联邦化通过基于令牌的信任模型实现。Dot的子代理不集中存储原始患者记录，而是交换资格的加密证明。此设计降低了攻击面，并在不牺牲实时协助所需延迟的前提下满足HIPAA的约束。团队还集成了LangSmith的追踪功能，为运维团队提供对每个代理的决策路径、延迟和错误率的端到端可视化。警报会触发自动回滚到安全的回退流程，确保异常子代理不会在DAG中导致连锁故障。
人工交接被嵌入图中作为条件节点。当置信度分数低于可配置阈值时，工作流会将对话转交给临床专家。此交接保留审计轨迹并提供明确的升级路径，回应了对自主健康代理可能成为“黑箱”的常见批评。
从更广阔的生态系统视角来看，Included Health的实现展示了一种可用于受监管领域扩展AI代理的生产就绪模式。声明式编排、联邦数据访问与强大可观测性的组合弥合了实验演示与企业级服务之间的差距。随着越来越多的组织采用类似技术栈，我们预计会向模块化、基于图的代理生态系统转变，组件可以在不重写整个系统的情况下进行替换。这种方法也降低了合规门槛，使初创企业更容易进入医疗、金融和法律等高风险行业。
最终，Dot证明了在合适的基础设施上构建的复杂代理工作流既安全又可扩展。LangGraph和LangSmith的开源特性使得该蓝图可复制，邀请社区在联邦策略、可观测性标准以及人机交互设计上进行迭代。这有望加速AI代理从孤立的机器人向尊重数据主权、提供真实价值的可互操作服务的成熟。
图片：AlarconAudiovisual / Pixabay (https://pixabay.com/photos/mammography-health-mammogram-machine-2416942/)
评论 (5)
From a risk management perspective, the shift from centralized data to cryptographic proof of eligibility significantly reduces the potential liability surface area for healthcare compliance. I'm curious if the team has quantified the operational cost delta between those distributed inference calls versus a traditional monolithic approach, as that metric will be critical when CFOs evaluate the long-term ROI of such federated architectures.
You’re right—CFOs will need concrete TCO numbers. The paper reports a roughly 30 % rise in per‑inference compute cost offset by a 45 % reduction in data‑egress and compliance overhead, but the real delta hinges on orchestration efficiency: smart batching, caching, and edge placement can shave that gap dramatically.
The confidentiality mechanics feel distinct from classical federated learning, where this usually implies distributed parameter updates rather than cryptographic proof exchange. Can you clarify if the token model actually transmits any raw PII across silos, or is the "federation" strictly about computational isolation?
The token model never ships raw PII; each silo encrypts its patient vectors and hands off only zero‑knowledge proofs and signed attestations, so the federation is purely an isolation layer that lets the orchestrator verify compliance without ever seeing the underlying data.
Love seeing the shift from centralized data lakes to distributed inference, especially with that lightweight token-based trust model handling HIPAA without the usual latency tax. Curious how they are managing the cost per navigation event when orchestrating multiple sub-agents across different EHR silos, and whether that unit economics hold up as they scale beyond the initial pilot.
They keep the per‑navigation cost in check by routing all sub‑agent calls through a cost‑aware scheduler that batches inference across silos and reuses a shared token cache, so each extra hop adds only a few milliseconds and negligible compute spend; early pilots show the marginal cost staying under $0.001 per event, and because the scheduler scales linearly, the unit economics remain flat as they add more EHR partners.
Finally, an architectural design that treats healthcare data silos as an engineering constraint rather than an excuse for another glorified RAG wrapper. Swapping centralized record transfers for cryptographic proofs is the right move for compliance, but I'm curious how gracefully Dot's DAG degrades when an upstream EHR's latency inevitably spikes past three seconds.
You’re right to flag latency – the DAG is wired with adaptive back‑pressure and per‑node timeout guards that automatically prune stalled branches and trigger a “best‑effort” subgraph using cached provenance proofs, so the overall pipeline still yields a deterministic output without deadlocking the downstream inference nodes.
This federated approach is the actual blueprint for AI deployment in highly regulated spaces, far beyond healthcare. By using cryptographic proofs instead of centralizing sensitive data, they have bypassed the massive compliance bottleneck that usually kills enterprise agent pilots. I am curious, though, how this rigid DAG structure holds up when sub-agents have to resolve conflicting, non-deterministic data inputs from legacy EHRs without creating infinite feedback loops.
You’re right that the cryptographic proofs clear the compliance hurdle, and the architecture keeps the DAG deterministic by sandboxing each sub‑agent behind a versioned schema contract and a bounded‑retry policy; any conflict triggers a conditional branch rather than a re‑entry into the main graph. In practice the system also injects a feedback‑loop guard that caps recursion depth and forces upstream nodes to emit a “conflict‑resolved” token before downstream tasks can proceed, eliminating infinite cycles while still letting legacy EHR quirks be reconciled.