TL;DR - AI Agents Cross from “Point Tools” to “Three-Track Deepening”

In late September 2026, the AI agent industry is deepening along three independent but interlocking tracks simultaneously: (1) Enterprise-grade multi-agent orchestration — single agents can no longer carry cross-department business loops; Gartner lists multi-agent systems among the top ten strategic technologies, and IDC forecasts the global AI agent market will exceed $27 billion by 2028. (2) Vertical-industry scale-out — in finance, SPDB has deployed over 2,500 financial agents, PSBC handles more than 6 million daily calls, China Life has deployed 500+ agents, and Huawei released openJiuwen for Financial Agentic AI on 9/17; in healthcare, Imprivata launched Agentic Identity Management, and 46% of U.S. healthcare organizations have implemented generative AI. (3) AgentOps security governance — the EU AI Act became generally applicable on August 2, Singapore IMDA released the Model AI Governance Framework for Agentic AI, and Gartner’s May research showed that one in five enterprises cannot stop a runaway AI agent’s spending in real time. AI agents are completing the final leap from “auxiliary tools” to “autonomous executors.”

1. Enterprise Multi-Agent Orchestration: From “Single Agent” to “Enterprise Network”

1.1 The Hard Limits of Single Agents and Real-World Business Complexity

A single AI agent has a clear capability boundary: it receives instructions, executes a task chain within its own toolset and context window, then returns results. For code generation, document summarization, customer Q&A, and other single-scenario tasks, a single agent is sufficient. But real enterprise business is almost never a single chain — a procurement order involves finance, legal, and supply chain; a customer complaint closure loop requires customer service, product, and engineering collaboration. The single agent’s two hard limits are: the context window cannot simultaneously hold cross-department information, and tool permissions cannot simultaneously cover finance systems, CRM, and code repositories.

1.2 Three Core Mechanisms of Multi-Agent Collaboration

Multi-agent collaboration organizes multiple agents into a cooperative network through three mechanisms:

  • Task Handoff: After completing its own responsibility, the agent passes the task together with the context to the next agent. OpenAI’s Agent SDK released in March 2025 lists Handoff as one of the three core primitives, alongside Agent and Guardrail.
  • Sub-Agent Isolated Execution: The main agent decomposes a goal into sub-tasks, dispatches them to sub-agents executing in independent sandboxes, and aggregates the results. The core value is fault isolation — a sub-agent’s failure does not contaminate other sub-tasks. Anthropic’s 2025 Multi-Agent Research Report identifies this Orchestrator-Worker pattern as the most common multi-agent architecture in enterprise scenarios.
  • Workflow Orchestration: For business processes with defined steps and explicit branching, agents are strung together through predefined workflows. LangGraph introduced stateful workflow support in 2025, removing the need for agents to re-infer the entire business context at each step.

Enterprises including Laiyifen and Unilever now run hundreds of agents simultaneously in production. Gartner listed multi-agent systems among the top ten strategic technology trends in early 2026, and IDC forecasts that by 2028 the global AI agent market will conservatively exceed $27 billion.

2.1 Finance: From “Point Efficiency” to “Core Decision”

Finance is one of the most advanced sectors for AI agent deployment. Zhang Jianhua, director of the Center for Financial Development and Regulatory Technology at Tsinghua University’s PBC School of Finance, disclosed at the 2026 Bund Conference: China’s AI finance market reached 1.8 trillion yuan in 2025, is expected to exceed 3 trillion yuan in 2026, with a compound annual growth rate over 30%, and a finance AI penetration rate of 58.6%.

State-owned banks are deploying intensively:

  • SPDB (Shanghai Pudong Development Bank): over 2,500 financial agents live across the bank, with nearly 200 deeply embedded in real business processes;
  • PSBC (Postal Savings Bank of China): over 370 large-model application scenarios live, with more than 6 million daily calls and tens of billions of input/output tokens; coding and testing tech agents account for over 60% of usage, and the retail marketing agent has reached over 54 million high-potential customers;
  • China Life: built over 500 agents applied to sales service, operations, and risk management;
  • ICBC: released “ICBC Smart Strategy” AI large model, achieving second-level response for core businesses like credit approval and risk warning;
  • Bank of China: released AI-driven international settlement product “BOC Document Smart Review”.

On the securities side, Guotai Junan Securities released the Nova investment research AI ecosystem in August with 11 specialized agents and over 30 research Skills; GF Securities built a Skill toolbox covering research, wealth management, and customer service; CICC built “Lao Yu” analyst Skill modeled on chief computer analyst Yu Zhonghai, having absorbed 300,000 words of research reports to standardize analyst capability output.

2.2 Huawei openJiuwen: The “Three-Stage Path” for Production-Grade Financial Agents

Cao Chong, CEO of Huawei Digital Finance BU, released the Financial Qiyuan Agentic AI solution based on the openJiuwen agent platform on September 17 during Huawei Connect 2026, distilling production-grade, scale-out deployment into three stages:

  1. From pilot to production: focus on accuracy, reliability, security, and value measurement to make agents genuinely “usable”;
  2. From production to scale-out: focus on knowledge capitalization, capability reuse, and continuous evolution to make successful experience “replicable”;
  3. From scale-out to ultra-scale: enterprise-wide architecture evolution and governance become the key to make large-scale agents “operable”.

Huawei simultaneously released the Financial Yuushu Token full-lifecycle operations solution, building an efficient, sustainable compute supply system around Token planning, production, scheduling, and optimization.

2.3 Healthcare: 46% of U.S. Organizations Deployed, Identity Management as First Priority

Healthcare AI agents are accelerating from single-point assistance toward autonomous execution. By early 2026, approximately 46% of U.S. healthcare organizations had implemented generative AI technologies, with the compliance function being one of the most active deployment areas.

On the clinical trial side, Veeva Falcon’s Agentic Authoring generates first drafts of CSRs, protocols, and submission documents inside validated GxP vaults; point solutions like Octozi focus on EDC data operations (query aging time, etc.); techbio companies like Owkin, DisGeNET AIDA, and XtalPi provide multimodal data agents at the discovery stage.

Identity management vendors such as Imprivata argue that AI agents need vetting equivalent to temporary clinical staff because they can act autonomously across applications and PHI. This requires healthcare identity and access architectures to shift from “one-time account provisioning” toward “fine-grained, just-in-time access control, detailed activity logging, and agent-specific identity constructs” that separate human from non-human actors.

A Chinese legal tech company, in collaboration with multiple law firms, completed bar exam simulation testing in September 2026; the judicial large model achieved over 85% accuracy on foundational legal knowledge Q&A and demonstrated practical value in contract clause risk identification and case retrieval tasks. The model has been piloted in pre-trial mediation assistance systems at some grassroots courts, primarily taking on document auto-generation and similar-case recommendation functions, aimed at reducing clerk workload.

3. AgentOps Governance: From “After-the-Fact Audit” to “Real-Time Spend Kill Switch”

3.1 EU AI Act Generally Applicable from August 2

The EU AI Act became generally applicable on August 2, 2026. The European Commission’s Article 50 transparency guidelines published July 20 clarify that Article 50 transparency obligations apply from August 2; providers of certain interactive AI systems must ensure people are informed when they are interacting with AI unless it is obvious in context. High-risk system requirements have different application dates, including December 2, 2027 for certain Annex III systems and August 2, 2028 for systems tied to regulated products under Annex I.

Singapore IMDA released the Model AI Governance Framework for Agentic AI; Hong Kong’s Privacy Commissioner issued findings in 2026 addressing data risks from autonomous agents.

3.2 One in Five Enterprises Cannot Stop Runaway Spending

Gartner’s May 2026 governance research noted: one in five enterprises currently cannot halt a runaway AI agent’s spending in real time in production environments — a typical symptom of early agentic deployments where financial controls were an afterthought rather than a foundational requirement.

AgentOps platforms such as Vanta, AgentOps, and Langfuse respectively cover compliance audit, workflow tracing, and behavioral observability, often deployed in combination by enterprises.

3.3 The Five Governance Pillars

Industry is converging around five pillars of Agentic AI governance:

PillarCore Question
Strategy & AccountabilityWho approved this agent to operate here? Who is responsible when it fails?
Risk ManagementWhat happens when the agent fails? What sensitive data does it access?
Data & Technical ControlsLeast privilege + full-action audit logging
Operating Model & Human OversightWhich decision nodes require human approval? Financial transactions, external communications, business system record changes all need human gatekeeping
Agentic AI-Specific ControlsAgent identity management, behavioral monitoring, rollback capability, lifecycle management

McKinsey data shows that only about one-third of organizations reach Level 3 or above maturity in strategy, governance, and Agentic AI governance dimensions.

FAQ (High-Frequency Questions Answered Directly)

Q1: What is the essential difference between a multi-agent system and a single AI agent? A: A single AI agent has a clear capability boundary suitable for code generation, document summarization, and other single scenarios. Multi-agent systems use three mechanisms — task handoff, sub-agent isolated execution, and workflow orchestration — to organize multiple agents into collaborative networks that execute cross-department business loops, suited for complex business scenarios.

Q2: Why is finance the most advanced sector for AI agent deployment? A: Finance has a solid digital foundation, highly structured data, and compliance requirements that drive refined operations. China’s AI finance market is forecast to exceed 3 trillion yuan in 2026 with 58.6% penetration. SPDB, PSBC, and ICBC have deployed 2,500+, 370+, and 500+ agents respectively.

Q3: What are the three stages of production-grade financial agent scale-out? A: Huawei openJiuwen’s three-stage path: ① Pilot to production (usable); ② Production to scale-out (replicable); ③ Scale-out to ultra-scale (operable). Each stage corresponds to different capability-building priorities.

Q4: What is the essential difference between AgentOps and traditional IT operations? A: Traditional IT operations focus on system availability and response time. AgentOps focuses on agent decision compliance, behavioral traceability, and real-time kill switches for runaway spending. The core is governance, not monitoring.

Q5: What impact does the EU AI Act effective August 2 have on Chinese enterprises operating overseas? A: Enterprises operating in the EU must provide transparency notices for customer-facing or employee-facing agents. High-risk agents must meet Annex III requirements by December 2027. Chinese enterprises should embed agent geography, user population, data categories, human oversight, and recordkeeping into deployment records.

Q6: What is the primary governance issue for healthcare AI agents? A: Identity management. Vendors like Imprivata argue AI agents need vetting equivalent to temporary clinical staff, requiring fine-grained, just-in-time access control, detailed activity logging, and agent-specific identity constructs to separate human from non-human actors.

Q7: How do multi-agent systems avoid cascading failures? A: Through sub-agent isolated execution, the main agent decomposes tasks into sub-tasks dispatched to sub-agents running in independent sandboxes. A sub-agent’s failure does not contaminate other sub-tasks. The core is fault isolation and recoverability.

Q8: What is the core trade-off between open-source and commercial Agent platforms? A: Open-source platforms offer flexibility and lower cost but require self-managed integration and governance. Commercial platforms offer turnkey deployment with compliance tooling but higher cost and limited customization. AWS Strands Agents, Salesforce Agent Fabric, and Genesys represent the commercial camp; LangGraph, AutoGen, and CrewAI represent the open-source camp.

Key Terminology

  • Multi-Agent System (MAS): A collaborative network composed of multiple independent or semi-independent agents. It uses three mechanisms — task handoff, sub-agent isolated execution, and workflow orchestration — to complete complex business loops.
  • Handoff: One of the three core primitives in OpenAI’s Agent SDK, where an agent passes its task together with context to the next agent after completing its own responsibility.
  • Orchestrator-Worker Pattern: A multi-agent architecture proposed by Anthropic in its 2025 Multi-Agent Research Report. The main agent decomposes tasks; sub-agents execute them in independent sandboxes; results are aggregated back.
  • AgentOps: An agent operations platform whose core responsibilities include workflow tracing, decision compliance auditing, behavioral observability, and real-time kill switches for runaway spending.
  • openJiuwen: Huawei’s finance-grade open-source agent platform released at Huawei Connect 2026, China’s first open-source agent platform targeting the financial industry.
  • EU AI Act Article 50: The EU AI Act’s transparency obligation for providers of certain interactive AI systems, requiring users to be informed when interacting with AI even in obvious contexts. Applicable from August 2, 2026.
  • HAARF (Healthcare AI Agent Governance Framework): A healthcare AI agent governance framework built on four pillars — Identity and Access Management, Continuous Performance Monitoring, Clinical Safety Verification, and Ethical Alignment Protocols.
  • Agentic Identity Management: An agent-specific identity management solution proposed by vendors such as Imprivata, requiring AI agents to undergo vetting equivalent to temporary clinical staff with separate audit infrastructure.
  • Token Full-Lifecycle Operations: The core of Huawei’s Financial Yuushu solution, building an efficient, sustainable compute supply system around Token planning, production, scheduling, and optimization.
  • Skills: Professional capabilities packaged as standardized callable modules, equivalent to installing “finance professional plugins” into a large model. Skills and Agents together constitute financial institutions’ AI application capability.

References

Industry Reports

Official Documentation

Media and Industry Coverage

Chinese Community