Enterprise agentic workflows and multi-agent systems for investment strategy: JPMorgan argues that deployable investment-research agents require governed autonomy, not unconstrained automation

Overview

JPMorgan's Global Quantitative Strategy summarizes a September 2026 workshop delivered with Google on building enterprise agentic workflows for investment research. The report argues that the practical objective is governed, reviewable autonomy: agents should automate repeatable, data-intensive tasks while operating within approved tools, data sources, acceptance criteria and escalation paths.

Core views

The report distinguishes agentic AI from traditional automation, machine learning and generative AI. Fixed automation follows predefined code paths, machine learning identifies patterns, and generative AI expands natural-language interaction with unstructured information. Agentic AI instead manages the workflow itself: it can decompose an objective, select tools, execute actions, review intermediate work and iterate. JPMorgan stresses that this does not mean unsupervised behavior. In investment workflows, autonomy must be bounded by permissions, review points and escalation paths so that the process remains explainable and defensible. The economic case is selective rather than universal. The report argues that AI is better suited to scalable, repeatable work such as document classification, drafting and elements of coding, whereas people retain an advantage in high-stakes judgment, strategic work and creative tasks. Whether deployment is rational depends on token costs relative to human time, time to completion, rework, error rates, consistency and oversight. Financial research is presented as a strong use case because it is multi-step and tool-heavy, but its direct connection to investment decisions means reliability and traceability are essential. The workshop's first specialist agent used a live market-data tool to retrieve current stock-price inputs and perform calculations at query time. This bridges the gap between a model's training-date knowledge and an evolving market, replacing the manual process of sourcing data separately and pasting it into prompts. The report emphasizes that speed is not the only benefit: approved data-source restrictions and defined behavior for missing, stale or inconsistent data are needed to prevent plausible but unsupported outputs. The prototype used a pricing function that defaulted to 100 when a ticker was not found, while the report separately highlights guardrails for unavailable data as a way to reduce hallucination risk. The second specialist was a RAG analyst for private documents. Documents are chunked, converted into semantic embeddings and stored in a vector database; a user query is embedded similarly, and similarity search retrieves relevant passages for the model's context. The report says this can ground answers in approved internal policies, market outlooks or corporate documents, helping the agent remain aligned with house views and relevant constraints. It also stresses that RAG is reliable only when tested: teams should inspect retrieved passages behind representative answers. If retrieval is correct but the answer is weak, instructions and response constraints should be tightened; if retrieval misses relevant evidence, chunking, embeddings or search settings should be adjusted. The illustrated RAG agent uses a 30-second timeout to prevent endless searching for information absent from the database. Manager-led orchestration is the next step. A routing manager interprets the user's objective, plans subtasks, delegates them to specialized agents and consolidates their outputs. JPMorgan argues that narrower remits reduce confusion and can improve output quality, while routing routine tasks to cheaper, faster models and reserving higher-capability models for complex reasoning can control costs. Its value becomes greater as an organization scales from a few tools to dozens or hundreds of specialists, when people cannot reliably track every agent's remit and access rights. Model selection, routing rules, accuracy, latency and cost should therefore be tested and tuned for each workflow rather than standardized through a single configuration. The final demonstration combines these components into an autonomous research team governed by explicit roles and stopping conditions. A goal-refiner turns a broad request into a precise research brief. A research loop combines a RAG analyst, a compliance officer and a termination checker; the checker determines whether the output meets the acceptance bar or requires another iteration. The illustrated loop has a maximum of four iterations. A reporter then produces an executive memo using only the research-loop output and is instructed not to fill gaps with outside knowledge. This hierarchy is intended to control quality and token consumption while making responsibility and evidence flow visible. The report concludes that a production-grade "harness" separates a clever prototype from a deployable system. A harness exposes plans, intermediate steps, tool calls, sub-agent activity and resource use, allowing users to diagnose issues and intervene. JPMorgan identifies three practical tuning levers: inference settings such as temperature to reduce variability, richer instructions and few-shot examples to standardize outputs, and testing to decide when functions should remain tools versus become delegated sub-agents. Governance is characterized as an engineering requirement, not an afterthought: systems need explicit permissions, review and escalation gates, iteration and runtime limits, and inspectable evidence trails as they scale.

Analysis framework

The report follows a build-up from the definition and economics of agentic AI, to a live-data specialist, a RAG specialist, manager-led routing and a full research-team architecture. It uses workshop prototypes to explain how approved data access, retrieval grounding, testing, role separation, compliance review and loop limits can make automated research workflows reviewable and controllable.

Methodology notes

• Other

  Retrieval-Augmented Generation (RAG)

  The workflow retrieves relevant passages from an approved, vectorized document set and supplies them as context, aiming to ground answers in source evidence rather than unsupported model knowledge.

• Other

  Manager-led multi-agent orchestration

  A manager agent decomposes a request, routes subtasks to specialists, combines results and applies explicit quality, cost and termination controls.

Key data

• RAG-agent connection timeout30 secondsDesigned to prevent an agent from endlessly searching for information not contained in its database.
• Autonomous research-loop maximum iterations4The loop ends when the acceptance standard is met or the iteration cap is reached.
• Autonomous research-team stages3Goal refiner, iterative research loop and reporter.

Impact & implications

The report presents governed agentic systems as a way to automate repeatable research steps while preserving evidence traceability and human oversight. It suggests that organizations should tailor tools, models, prompts, routing and controls to specific workflows rather than deploy a generic autonomous agent.

Risks

• Hallucinations can propagate through an end-to-end automated workflow when an early error is not detected.
• Outputs may appear confident even when underlying data was not retrieved, is missing, stale or inconsistent.
• Unbounded loops, excessive tool calls and poorly defined agent permissions can create uncontrolled cost, safety and reliability problems.
• RAG outputs may be unreliable if retrieval quality, chunking, embeddings, search settings and source grounding are not tested.

What to watch

• Whether firms select workflow use cases where time and monetary costs are justified by measurable gains.
• Testing results for model choice, inference settings, instructions, tool configurations, accuracy, latency and cost.
• Whether production systems expose tool calls, intermediate outputs, evidence sources and resource usage for review.
• Whether agent workflows have explicit permissions, acceptance criteria, escalation paths and iteration limits.