Frost & Sullivan, in collaboration with LeadLeo, has released the '2026 China Enterprise-Level Agent Intelligent Agent Solution Market Report'.

The enterprise-level Agent intelligent solution is a full-stack technical system centered on enterprise data and business logic, divided into four layers: the model and reasoning layer consists of a base model, expert model, reasoning engine, and Token factory, providing cost-effective model capabilities per scenario; the data and knowledge layer aggregates the enterprise-level knowledge base, data foundation, business logic, and long-term memory; the construction and orchestration layer, centered around Agent Harness, is responsible for context management, tool invocation, and multi-agent orchestration; the delivery and application layer supports industry agents, functional agents, and digital employees. Tool connection and operation governance span all layers, ensuring the safe and controllable operation of the Agent.

The key difference between enterprise-level Agent and ordinary AI assistant lies in planning tasks around business objectives, invoking system actions for execution, and maintaining a complete evidence chain—its core value is the creation of a manageable, traceable, and iterative business execution architecture. Application depth increases from L1 to L5, and risk and governance requirements rise accordingly: L2-L3 focus on knowledge processing, material generation, and process embedding, allowing risks to be controlled and value to be easily verified; L4 is suitable for a few tasks with clear rules and verifiable results; L5 requires the system to operate in a closed loop without human intervention, demanding the highest level of governance capability. The report suggests that enterprises should prioritize accumulating replicable value at L2-L3 before advancing to supervised execution and autonomous closed loops.

Definition and research scope of the enterprise-level Agent intelligent solution solutions

Industry applications of enterprise-level Agent are generally at L2-L3. The business objects and feedback loops in internet technology, telecommunications, and e-commerce retail are highly online, forming closed loops in R&D, customer service, trading, and operation, enabling early entry into L3-L4; applications in industries such as finance, government, and healthcare have a broader range but mostly remain in the stages of knowledge processing, material generation, and process assistance—application breadth does not equate to autonomous depth; strong regulatory environments and physical industries, constrained by responsibility boundaries and physical verification, still primarily rely on controlled applications.

From a functional scenario perspective, enterprise-level Agent exhibit patterns such as leading in customer support, accelerating process coordination, and promoting cautious advancement in key decisions. Customer support features standard knowledge, clear work orders, accessible systems, and stable service metrics, enabling early achievement of L3; operations, engineering R&D, marketing, human resources management, sales, and finance are concentrated in L2-L3, as the aspects involving capital, personnel, and professional judgment still require manual approval and responsibility tracing. The maturity differences among similar functions across different enterprises stem primarily from their own processes and data foundations.

The primary reason for enterprises to deploy Agents is to unlock the collaborative value of existing digital systems. Most enterprises have established knowledge bases, CRM, ERP, work order, and data platforms, but business tasks still rely on employees to cross-system queries, transfers, entries, and follow-ups; the Agent connects dispersed systems through business objectives, and its recent value lies in shortening business cycles, reducing process breakpoints, and enhancing execution consistency—direct replacement for full positions remains limited. The report indicates that functional value should be measured by business outcomes; only by continuously improving indicators such as problem-solving rate, processing time, incremental conversion, and closing cycle, can projects move from local pilot programs to cross-departmental replication.

General overview of the penetration depth of enterprise-level Agent intelligent solutions across industries

Enterprise-level Agent technology has gone through four stages: dialogue generation, retrieval augmentation, workflow intelligence, and full-stack Agent system. The first two stages address accuracy, while the latter two address reliability; an Agent’s errors no longer limit themselves to text but transform into actual business losses, leading to the development of context engineering, tool protocols, memory, and operation governance as independent technical layers. The evaluation criteria for technical solutions shift from answering accuracy to task completion rate, operational stability, and single-task cost.

At the model layer, supply is shifting from a single general large model to a combined supply of base models, expert models, and reasoning engines, where the unit Token cost has become a key factor restricting large-scale deployment; at the data and knowledge layer, the performance of the same model across different enterprises primarily depends on the organization quality of enterprise knowledge and data; at the construction and orchestration layer, Agent Harness manages context, tools, memory, and execution loops, and the same base model paired with different Harnesss may yield significantly different task completion outcomes. Open protocols such as MCP and A2A, after reducing connection costs, will show greater differences in ability to locate, retry, and rollback when long-chain tasks fail.

The operation and governance layer determines whether the Agent can transition from a pilot to a production environment. Once the Agent gains system operation permissions, security risks expand from content risks to target hijacking, tool abuse, and permission abuse; the gap between demonstration and production is primarily reflected in three capabilities: manageability, observability, and controllability. The report recommends that enterprises embed identity, permission, audit, and manual approval into runtime prior to launch and establish interception and rollback mechanisms for high-risk actions.

Core functions of Agent Harness

The growth drivers for enterprise-level Agent have shifted from technology supply to joint promotion from both supply and demand. The 'AI +' initiative clarifies the popularization direction for intelligent agents; reduced reasoning costs and widespread adoption of open protocols lower the deployment barrier, creating sustained demand for process efficiency and business outcomes; however, in industries with strong regulations or high responsibilities, such as finance, government, healthcare, and industrial, compliance requirements and responsibility boundaries will still determine the implementation pace. The bottleneck of large-scale deployment has extended from model capabilities to enterprise process governance: if process breakpoints are not identified, data consistency is not ensured, permission boundaries are unclear, and the responsibility chain is not closed, even a well-performing Agent during the demonstration stage will struggle to sustain operation in the real business environment.

The report states that enterprise-level Agent will develop along two main lines: product architecture upgrades and business model deepens. The product form will evolve from personal efficiency tools to a process execution system, with multi-Agent collaboration, industry knowledge enhancement, AgentOps governance, and enterprise software native integration becoming key architectural elements; the business model will shift from tool procurement to value delivery centered on scenario implementation, continuous operation, and outcome validation, with project pricing moving from seats or calls to scenario delivery and business outcome validation, and evaluation focusing on task success rate, process cycle, cost, stability, and risk indicators.

Product and architecture trends for enterprise-level Agent