The central challenge for enterprise AI is shifting. While model capabilities continue to improve, the bottleneck hindering AI commercialisation has moved from "is it usable?" to "can AI be stably embedded into core business processes and consistently deliver verifiable results?" In this context, the collaborative model between the Frontline Deployment Engineer (FDE) and the Agent runtime system, Harness, is becoming a crucial pathway for scaling enterprise AI.

Analyst Yang Lin from Guotai Junan Securities' computer industry team highlighted in a report on August 13 that FDEs are responsible for deep engagement with business scenarios and accumulating industry expertise, while Harness ensures the stable execution and reusability of agent capabilities. Their synergy is expected to enable more efficient and replicable large-scale AI deployment for enterprises.

This assessment points to a clear shift in the value chain – moving from "who has the most powerful model" to "who better understands the business, who can deliver a production system, and who can replicate the experience from one project to the next." For the market, this implies a re-evaluation of the valuation logic for companies that possess only model or API access capabilities.

Investment focus is also adjusting accordingly. The two types of companies truly worth tracking are: those that can deeply integrate into industry-specific environments, embedding AI into clients' core processes; and platform providers that possess foundational elements like Agent Runtime, AIOS, and Ontology, allowing them to transform delivery experience into reusable software assets.

A significant gap exists between penetration and scale.

High AI application penetration does not equate to mature AI commercialisation.

McKinsey's "The State of AI 2025" report reveals a clear divergence. While nearly 90% of surveyed companies have normalised AI use in at least one business function, and 62% are experimenting with AI Agents, only about one-third have reached the full enterprise-scale stage. Even among large enterprises with revenues exceeding $5 billion, less than half have completed large-scale deployment.

The realisation of business value is also limited. Although many companies have seen cost savings or revenue increases in specific areas, only 39% of respondents reported that AI has impacted their overall EBIT, and for most, AI's contribution to EBIT remains below 5%.

The root cause lies in insufficient process restructuring. Among the 25 factors McKinsey analysed for impacting generative AI value realisation, redesigning business workflows showed the highest correlation with EBIT improvement. However, only 21% of companies using generative AI have fundamentally redesigned at least part of their workflows. A common characteristic of high-performing companies is that they are not just adding an AI tool, but rewriting processes, data, permissions, human-machine division of labour, and KPIs around AI.

When Agents enter the production environment, risks are amplified. The survey shows that 51% of companies using AI have experienced at least one negative consequence, with nearly one-third encountering problems caused by inaccurate AI output. While Copilot-like assistant tools can rely on human oversight, Agents need to call tools, interact with operating systems, and execute tasks across multiple steps. Errors accumulate, state drift occurs, and permission misuse can all escalate into production risks.

The core value of FDE lies in transforming on-site experience into reusable assets.

An FDE is not a traditional pre-sales role, nor a general implementation engineer. Traditional pre-sales addresses "why the customer should buy," the implementation team handles "how to deploy the product," and consulting firms answer "what the customer should do." An FDE needs to answer three questions simultaneously: which business problems are worth solving with AI, how to actually build the AI solution, and whether the business metrics have improved after launch.

This role consistently operates between the client's business front-line and product R&D, handling scenario identification, proof-of-concept, system integration, production deployment, and effect evaluation. Their output has a dual nature: for the client, it's a business outcome; for the vendor, it's a reusable asset.

The key differentiator between the FDE model and an outsourcing model lies in how experience is accumulated. If a project concludes leaving only a customised system for the client, it's essentially outsourcing. If it brings back experience but cannot be reused, it remains a project-based model. The FDE model's premise is to transform on-site experience into Skills, connectors, industry templates, test sets, workflows, Ontology, or platform capabilities, thereby continuously reducing the delivery cost for similar clients. The true metric for measuring FDE scalability is whether similar clients can be onboarded faster, with fewer FDE person-days, a higher proportion of reusable assets, and ultimately, client revenue growth outpacing the growth rate of the delivery team size.

AI coding and Agent tools are reshaping the economics of deep customisation. Repetitive tasks involving code, interfaces, test cases, and documentation can be partially automated, allowing FDEs to dedicate more time to requirement clarification, architectural decisions, and abstracting industry knowledge. This is the underlying logic behind leading players like OpenAI and AWS accelerating their FDE organisation build-out. Reports indicate that OpenAI has established a Deployment Company and plans to acquire Tomoro, an enterprise AI consulting and engineering firm, bringing in approximately 150 experienced FDEs and Deployment Specialists. It is also collaborating with 19 investment firms, consultancies, and system integrators, committing over $4 billion to expand enterprise AI deployment capabilities. Meanwhile, AWS has invested $1 billion to create a Forward Deployed Engineering organisation, aiming to compress Agent deployment cycles from months to days.

Harness determines whether an Agent can operate sustainably in a production environment.

Harness is the Agent runtime system that wraps around the base model. The model handles understanding and reasoning, while Harness manages context, memory, tool invocation, state maintenance, permission control, result validation, anomaly recovery, and long-running task management. It dictates what the model can see, what it can invoke, when it should act, and how to recover from errors.

Increased task complexity exponentially amplifies the importance of Harness. A simple chatbot only needs input, generation, and output. A production-grade Agent requires planning, retrieval, tool calls, receiving environmental feedback, and re-decision-making, sometimes running continuously for hours. The longer the task chain, the more easily a single point of failure can evolve into a systemic risk.

Anthropic's experimental data vividly illustrates the cost and value of Harness. Using the same Opus 4.5 model to develop an application, a single Agent took about 20 minutes at a cost of roughly $9. A complete Harness setup, including Planner, Generator, and Evaluator, ran for about six hours at a cost of around $200 – a cost over 20 times higher. However, the application's completeness, usability, and correctness of core functions were significantly improved. This shifts the competitive metric from "cost per token" to "Cost per Successful Task" – the total cost for successfully completing a task each time.

Microsoft has productised Harness, integrating capabilities like Function Invocation, History Persistence, Context Compaction, Tool Approval, and Telemetry into a unified runtime. Developers can obtain a relatively complete Agent Runtime by configuring the model, instructions, and tools. It is noteworthy that the competitive focus of Harness will not remain on the number of components. As model capabilities continue to evolve, some task decomposition and context processing within Harness can be gradually handed back to the model itself. However, components like the Planner and Evaluator still hold irreplaceable value for tasks that exceed the model's reliable boundaries. Anthropic's view is that Harness should be continuously tested, pruned, and adjusted in response to changing model capabilities, rather than being a fixed process.

FDE and Harness work together to form a closed-loop for enterprise AI commercialisation.

FDE and Harness solve problems at different levels, but only through their synergy can a sustainable commercialisation loop be formed.

The logical chain is as follows: The FDE enters the client's environment to acquire genuine business knowledge – process structures, permission settings, types of exceptions requiring human intervention, results needing explainability, and metrics for proving ROI. Harness encodes this knowledge into executable Agent capabilities: Skills, Tools, Workflows, Ontology, permission rules, runtime constraints, and evaluation sets.

After launch, the production environment continuously generates real feedback – tool call failures, state loss, process anomalies, edge cases, and records of human takeover. The FDE diagnoses the root cause of failures and re-deposits the new experience back into the Harness and platform. This creates a flywheel: on-site knowledge enters the system, system operation generates feedback, and feedback is abstracted into reusable assets. As similar clients accumulate, new projects shift from re-development to configuration with minimal customisation. The competition ultimately becomes not about who can dispatch more engineers, but who can transform human on-site experience into machine capabilities faster.

Overseas vendors have already run this path in different forms. Palantir, with its Ontology at its core, has FDEs map business objects, data structures, and decision constraints to an enterprise operations layer containing Data, Logic, Action, and Security, with AIP providing end-to-end observability. Microsoft deeply integrates FDEs with industry partners, jointly investing over $1 billion with EY to deploy Agentic AI across core functions like Finance, Tax, Risk, HR, and Supply Chain. EY, as Client Zero, first validated the technology within its own organisation, with Copilot initially covering 150,000 employees and yielding a productivity improvement of approximately 15%. AWS has formed a two-layer Harness architecture: the AgentCore Harness ensures the stable operation of a single Agent, while the Delivery Harness accumulates Domain Ontology, Evaluation Framework, and Context Graph, enabling partners to independently replicate deliveries.

Domestic vendors show differentiated paths, with industry on-site experience and governance infrastructure as core differentiators.

Domestic vendors have established differentiated strategies in the FDE and Harness directions, with the primary distinction lying in the depth of their industry on-site experience and the completeness of their governance infrastructure.

Tencent Cloud's ADP has evolved from RAG and Workflow to natural language construction and introduced a cloud-based Agent Harness, supporting 24/7 centralised operation, a cloud Sandbox, and long-running tasks. Its CodeBuddy/WorkBuddy customer success team covers the entire chain from management awareness to business process alignment and effect evaluation. Some key clients are already purchasing FDE accompaniment services separately from tool procurement. Alibaba Cloud has officially started recruiting for "Front Deployment Engineers (FDE)," with its Bailian platform supporting both Agent and Workflow development modes. Its Wukong product is positioned as an enterprise-level AI-native work platform, pushing Agents into complex, cross-system tasks.

Kingdee International has embedded Harness within its ERP core. Its Lingji AIOS integrates the Harness system within the OS kernel, offering multi-agent orchestration, a Skill Hub, MCP gateway, and AI effect evaluation. The core is to encapsulate enterprise business context, permissions, workflows, and governance capabilities. Saiyi Information deploys FDEs in industrial environments, using its SDD standard development platform and a unified DevOps base to consolidate delivery capabilities, achieving a delivery cycle compression of up to 36% and an average reduction in manpower costs of over 16%.

iSoftStone has proposed an "AI Factory + FDE" model, connecting AI consulting, data intelligence, Agent development, and large model deployment and tuning. iFLYTEK, in a smart procurement project for the National Energy Group, transformed expert experience into 12 categories of standardised knowledge bases. The smart review system for procurement projects has stably operated for over 210,000 projects, achieving a 97% accuracy rate in smart reviews. Yixin Group uses Harness for financial Agent governance. As of the end of May 2026, its AI platform had been invoked over 125 million times, with the proportion of business completed autonomously by AI agents reaching 65%.