JP Morgan刚刚把内部多智能体系统Ask David的完整架构公开了。 个人觉得在很多场场景有参考学习的意义，构建多Agwnt框架可以使用。 这套系统在投资研究领域已经跑通，核心模式和当前最火的Agent架构高度一致： - Supervisor agent负责整体编排 - 专业subagent分别处理检索、结构化数据、分析等细分任务 - LLM-as-judge作为反射节点，在最终输出前做质量把关 - Human-in-the-loop填补最后一道准确性缺口 最值得注意的是，这套模式正在多个领域反复出现。 它证明了：真正能落地的多智能体系统，不是简单堆模型，而是清晰的分工 + 监督 + 反思 + 人工兜底的闭环架构。 对所有在做Agent的人来说，这段视频值得反复看。 你觉得Ask David这种架构，会成为企业级Agent的标准模板吗？
Conference-slide diagram of J.P. Morgan’s Ask D.A.V.I.D. architecture showing the supervisor agent (orchestrator), specialized sub-agents for structured data, unstructured (RAG) retrieval and analytics, a reflection node (LLM-as-judge), and human-in-the-loop — directly illustrating the multi-agent orchestration, evaluation/quality-check, and human fallback described in your summary.
Source: CameronRohn.com (LangChain // Interrupt 2025 photos & slides)
Research Brief
What our analysis found
JP Morgan Chase's Private Bank has publicly detailed the architecture of its multi-agent AI system, "Ask David" — short for Data Analytics Visualization Insights and Decision-making assistant — which was presented at the Interrupt conference in 2025. The system was built to automate investment research processes that previously required manual database searches and analysis, managing thousands of investment products backed by decades of data and handling billions of dollars in client assets. According to the team, Ask David has delivered striking efficiency gains, reducing research task time by up to 95% and targeting a 50% increase in the number of clients advisors can manage over a three-to-five-year horizon.
The system is built on a multi-agent architecture using LangGraph, featuring a supervisor orchestrator that delegates tasks to specialized sub-agents for structured data queries, unstructured data retrieval via RAG, and analytics and visualization. A reflection node powered by an LLM-as-judge mechanism checks output quality before delivery, while human-in-the-loop oversight ensures a final layer of accuracy — the real "David" is still consulted when stakes are high. The architecture closely mirrors patterns gaining traction across the broader AI industry, where reasoning models now comprise over 50% of AI usage and agentic workflows are becoming the dominant paradigm in enterprise AI.
Ask David is part of JP Morgan's wider AI strategy, which includes tools like LLM Suite and Coach AI and has contributed to an estimated $1.5 billion in cost savings across fraud prevention, personalization, and trading analytics. Multiple industry observers have described the system as a potential blueprint for enterprise-level multi-agent deployments, though the team acknowledges that achieving 100% accuracy with AI alone remains an ongoing challenge in high-stakes financial environments.
Fact Check
Evidence from both sides
Supporting Evidence
Multi-agent architecture confirmed
Ask David is explicitly described as a multi-agent AI system built on LangGraph, featuring a supervisor orchestrator that interprets user intent, maintains memory, and delegates tasks to specialized sub-agents — directly matching the tweet's claim of a supervisor agent responsible for overall orchestration.
Specialized sub-agents for distinct tasks
The system deploys a Structured Data Agent (translating natural language to SQL/API calls), an Unstructured Data RAG Agent (searching emails, reports, meeting notes), and an Analytics and Visualization Agent — confirming the tweet's description of sub-agents handling retrieval, structured data, and analysis.
LLM-as-judge reflection node verified
JP Morgan's architecture incorporates a reflection node using an LLM judge to assess output accuracy and retry if necessary, aligning precisely with the tweet's claim of an LLM-as-judge quality control mechanism before final output.
Human-in-the-loop oversight is integral
The system explicitly includes human-in-the-loop oversight, with the team noting that "Ask David still consults with real David whenever needed" for high-stakes financial decisions, confirming the tweet's claim about human oversight filling the final accuracy gap.
Proven in investment research with major efficiency gains
The system reduces research task time by up to 95% and targets a 50% increase in client capacity over three to five years, supporting the tweet's assertion that the system has been successfully deployed in investment research.
Emerging as an enterprise blueprint
Multiple sources describe Ask David's architecture as a replicable model for enterprise-level multi-agent systems, with principles applicable beyond finance — supporting the tweet's suggestion that this pattern is appearing across multiple domains.
Contradicting Evidence
100% AI accuracy remains elusive
The JP Morgan team openly acknowledges that achieving full accuracy with AI alone in high-stakes financial applications is not yet possible, which tempers the tweet's implication that the architecture has fully "run through" investment research without qualification. The system is designed to augment, not replace, human expertise.
Enterprise AI scaling faces systemic challenges
Broader industry analysis shows that scaling multi-agent AI systems across real production environments often encounters fragmented data ecosystems, inconsistent data standards, and governance hurdles — suggesting that replicating Ask David's success as a "standard template" across enterprises may be significantly harder than the tweet implies.
AI output is interpretation, not precision engineering
Experts note that AI-generated outputs represent interpretations rather than exact results, meaning that managing client expectations and handling revisions adds complexity when deploying these systems at enterprise scale — a nuance absent from the tweet's framing.
Architecture was not fully "open-sourced"
While the tweet states JP Morgan "公开了完整架构" (publicly released the complete architecture), the information was shared through a conference presentation and subsequent coverage rather than as a formal open-source release of code or detailed technical documentation, which is a meaningful distinction for practitioners hoping to directly replicate the system.
Report an Issue
Found something wrong with this article? Let us know and we'll look into it.