Enterprise Agents are about to enter a period of explosive application growth. A large number of enterprises have the opportunity to use AI to quickly catch up on digitalization at low cost, gradually moving towards being AI-native. "Claude Code pursues user experience more extremely in its Harness architecture for standalone and coding scenarios," said Lai Gangbin, CEO of Gemii Technology. In his view, Harness will not be swallowed by models; the two are mutually reinforcing. Harness is more like a new operating system in the AI era. He understands the future competition of Agent platforms as a series of technical route choices: how to bet on model capabilities versus Harness, whether to go local or cloud, and whether to build workflows or general-purpose Agents. For small and medium-sized tech companies, betting on the wrong route for three to six months could mean elimination from the market. During the interview, Lai Gangbin repeatedly emphasized that context is the only window for interacting with models. Self-evolution, memory, skills, and connectors are essentially sub-problems of context. Over the past year, Gemii Technology has evolved from an enterprise service software and operations provider to an enterprise intelligent agent platform. It chose the route of self-built enterprise intelligent agent platforms: deeply integrating enterprise context, knowledge bases, and application plugins into the platform, deployed entirely on the cloud, thus making it easier to connect to existing enterprise business containers. Many enterprises choose major vendor platforms, while others choose to build their own. How to be more out-of-the-box than open-source frameworks and closer to business scenarios than major vendor platforms has become the unique development path for Gemii's intelligent agent platform. In the previous decade, Gemii was a leading private domain marketing software and technology operations provider nationwide, serving many top enterprises in retail, consumer goods, healthcare, etc.; after the new enterprise intelligent agent foundation started running, it migrated traditional private domain businesses upwards while achieving breakthroughs and growth in self-built enterprise intelligent agent platforms, knowledge bases, and digital employee solutions.
Core Viewpoints
- The main battlefield for this round of enterprise AI is catching up on digitalization.
Chinese enterprises' digital construction lags behind the United States and is using AI to supplement infrastructure at extremely low cost. Currently, domestic enterprises with genuine data governance and knowledge bases are rare.
- Foresight in technical routes determines survival.
This industry competes not on team size, but on the core team's judgment of future technical routes.
- Context engineering is the core anchor of Harness.
The only output when interacting with models is context; self-evolution, memory, skills, connectors, etc., are essentially sub-problems of context.
- There are three paths for building context: Harness engineering, model post-training, and FDE manpower.
All three are different engineering implementation means for the same goal; the specific choice depends entirely on ROI.
- The organizational form of Chinese FDEs is more suitable for OPC (One-Person Company).
The supply of FDEs in the Chinese market is extremely scarce; the real opportunity lies in small organizations that understand both business and Agents.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01、What Chinese Enterprises Are Doing Now Is Using AI to Catch Up on 20 Years of Lessons
Analyst: Have you tried any new AI products recently? Have you tried Jev models?
Lai Gangbin: I've basically tried them all. What I use most daily is still Claude and our own intelligent agent products. Decision-making model products like Jev still need observation, especially comparing whether "small models + agent hooks" truly offer better value for money.
Analyst: Models and consumer-grade products change rapidly, yet enterprise Agent platforms seem stagnant. Why is there such a difference?
Lai Gangbin: I strongly agree with a saying in the industry: Without context, there is no way for Agents to deliver business value. The first step for 80-90% of Chinese enterprises now is to use AI to complete digital construction at low cost. AI provides a great opportunity to quickly realize the digital infrastructure that should have been built in the past ten years at one-tenth the cost compared to the SaaS era.
Analyst: On one hand, the market believes enterprise applications need to be reconstructed with AI; on the other hand, reconstruction requires waiting for underlying infrastructure to mature. Isn't this a contradiction?
Lai Gangbin: I think technology is no longer the issue. In the past decade, digitalization wasn't a technical problem, and neither is AI now. The key is whether enterprises can truly utilize their data—for example, how to actually bring what's under the iceberg to the surface. AI can help with better data cleaning and governance, structuring knowledge bases and documents effectively. We've talked to many peers, including teams from FanRuan and Alibaba, and the common sentiment is this: From the second half of last year to now, the things we can truly accomplish—making AI deliver real value—still start with data governance, knowledge bases, and structured cleaning of documents. This isn't just about new data. There are massive amounts of scattered materials in enterprise network drives and databases, tens of GBs or hundreds of GBs of historical data. How to effectively govern this before handing it to Agents remains the core task at this stage.
Analyst: After completing this layer of governance, how significant is the improvement in upper-layer application effectiveness?
Lai Gangbin: I think it's quite good. Some AI-native enterprises move fast and don't need to deal with so much historical dirty data governance. They use AI for advertising placement, sales enablement, customer service, and e-commerce content generation, which are already booming. Many private enterprises in Zhejiang and Guangdong are among those with the highest per-capita token usage.
Analyst: An unavoidable part of enterprise AI is action—enabling Agents to write back to business systems. Which enterprises have already achieved this?
Lai Gangbin: Small and medium-sized organizations push faster because they have no historical baggage and fewer compliance restrictions. Enterprises with very strong data compliance and business system compliance requirements move very slowly—it's a comparison between day and quarter levels. The speed of progress is also highly correlated with the extensive SaaS infrastructure previously built. If SaaS doesn't transform forward, it will likely become tools called by Agents; I highly agree with this. Our own SCRM and MA tools are undergoing MCP-ification and application containerization. Some systems have faster API/MCP integration, allowing direct calls; others require Computer Use methods. Technically, this is gradually becoming achievable; the differences lie in opening efficiency and daily token costs.
02、Three Bets: Context, Cloud, and Model Neutrality
Analyst: When enterprises talk about Harness, their focus varies. Some focus on context, others on self-evolution, governance, or model costs. What do you consider the priority?
Lai Gangbin: These directions are consistent in essence; they all belong to the construction of a complete Harness, i.e., the framework and engineering system supporting Agent execution. This year, everyone is mainly designing and upgrading Harness at the application layer. In the future, model post-training will develop further, but at the current stage, most enterprises are still in the phase of data governance and knowledge base governance. Products like LlamaIndex and RAGFlow haven't solved the data problems under the iceberg; RAG-type products assume documents are already well-organized, but reality is often not so. Whoever can feed large amounts of contextual data to Agents will win the first-phase market.
Analyst: Why place context at the most core position?
Lai Gangbin: Harness ultimately produces a complete set of contexts. Contexts are handed over to large models for reasoning. Context is the only window for interacting with models, so Harness must treat context engineering as its core anchor. The form of Harness will continue to evolve, and the peripheral ecosystem will enrich like an operating system. Over the past few months, the entire ecosystem around Harness has been gradually enriching, which is exciting. Meanwhile, from the perspective of developing our intelligent agent platform, we still need an anchor, which we define as upgrading and optimizing around context engineering. Whether building plugin systems, application containers, or multi-Agent setups, everything eventually returns to context: how to optimize context caching, handle syntax adaptation with models, etc. Solving these tends to yield better results.
Analyst: What obvious changes has Harness undergone in the past year?
Lai Gangbin: This year has been very dense. Initially, discussions focused more on single-turn optimization, then session-level optimization, followed by multi-Agent and A2A. Its capability boundaries are expanding, making it increasingly like an operating system. Before July and August this year, the iteration curve of Harness engine logic was very steep. After July and August, changes in the trunk slowed slightly, while peripheral ecosystems like plugins, security, sandboxes, and computing infrastructure entered a rapid iteration period. New upgrades include cross-Harness A2A communication, enabling collaboration between Agents on different platforms. Skills, connectors, and MCP are also being further encapsulated as plugins. Underlying data, sandboxes, and computing infrastructure themselves have many areas needing iteration and optimization.
Analyst: What are the differences in Harness design between Claude Code and Codex?
Lai Gangbin: Claude Code is more aggressive and designed more around local-end user experience. Codex has a more complete and standardized architecture, with a more robust product experience, resembling big-company approaches. Early on, Claude Code's user experience was clearly better. We judged at the time that Codex would benefit from a more robust technical architecture and might catch up or even surpass in the medium term. Over the past six months, this judgment has been gradually validated. In turn-level design, we reference Claude Code more; in complete session-level design, we reference Codex more.
Analyst: What needs to be solved most when implementing multi-Agent systems?
Lai Gangbin: Mathematically, it's 0.99 to the power of N. If a single Agent has 99% reliability, chaining ten reduces it to ninety percent. Each Agent has uncertainty, and multi-Agent systems multiply this uncertainty cumulatively. Ultimately, the challenge is controlling probability and uncertainty. Long-horizon tasks themselves aren't the problem. Knowledge base tasks can run for hours or even a day; the key is whether the task is sufficiently deterministic and whether the peripheral Harness is monitorable and correctable.
Analyst: Regarding the technical routes for enterprise Agent platforms, which directions do you bet on?
Lai Gangbin: The first bet is context. The Manus team proposed context last September or October. We developed our entire Agent platform this year, heavily borrowing ideas from Manus, such as cloud-based deployment, sandboxes, and relentless optimization around context engineering. Because we bet on context, we didn't touch workflows initially. Workflows have too many constraints and aren't AI-native enough; the direction must be towards generality. The second bet is cloud. Many office products from big vendors have local clients, but Gemii is entirely cloud-based. If targeting office scenarios, it 100% requires local clients because a lot of daily work happens locally. But we haven't touched office work; we focus more on AI transformation around digital employees and business flows. Looking ten years back and ten years forward, the trend is definitely cloud-based business flows, so we bet on cloud. The third bet is model neutrality. Models will have layers and distinct characteristics, while cost reduction and iteration speed will be very fast. The mainstream view says Agents differ from SaaS because every Agent service call incurs token costs, making Agent business less profitable than SaaS. But one variable cannot be ignored: the entire model layer, including open-source and small models, iterates capabilities very fast, and costs are dropping rapidly. My computer runs Qwen's 9B model and Google Gemma's 4B, both latest open-source local small models. Working on a plane, calling these small models is sufficient. For simple scenarios like customer service, running a 27B open-source model works 90% of the time. When small models are powerful, it becomes SaaS logic rather than a pay-per-token logic, allowing margins to become very high. These three bets combined form Gemii's current product form: a self-built enterprise intelligent agent platform integrating enterprise context, knowledge bases, and application plugins, running entirely on the cloud, thus easier to connect to existing enterprise business containers.
Analyst: So, in the competition for enterprise AI platforms, foresight in technical judgment is the decisive factor?
Lai Gangbin: Yes. Missing three to six months could mean falling behind. This era competes on technical routes and foresight, and on intersection points with business scenarios. It has nothing to do with team size; it relates to the core team's key judgments.
03、How to Build Context: Three Parallel Paths
Analyst: Since context engineering is the anchor, how exactly is context built?
Lai Gangbin: There are several paths for building context. One is the Harness engineering layer, optimizing skills, plugins, connectors, memory, etc. Another is via the model layer, such as post-training weight adjustments, or injecting special marker regions for enterprises within turns. The third is the manpower layer, i.e., FDEs, responsible for transporting business scenarios and data to Agents. These paths point to the same goal; the specific choice depends on Token ROI measurement.
Analyst: In the phase where models haven't reached stability, how do enterprises decide whether to do post-training?
Lai Gangbin: Still based on ROI. For example, in AI-native businesses, some want to do comics or content generation; if training doesn't achieve the effect, you must train. It's not about always training or never training. Most enterprises don't care about tokens, model parameters, or inference intensity; they only care about cost and effectiveness. As long as a service passes business assessment, it can proceed.
Analyst: Context relies on data access. Is there a significant problem in this link during enterprise implementation?
Lai Gangbin: Except for foreign enterprises and industries like healthcare temporarily stuck on compliance, the technical threshold for data access in most industries has passed. Enterprises can obtain context via APIs, CLIs, MCP, or Computer Use; the differences are mostly in efficiency and cost. But platforms only provide infrastructure. Specific scenarios require clear definitions of what data to collect and how. If there are no clear APIs or MCPs, FDEs need to intervene in enterprise context governance. From understanding requirements and collecting data to writing business applications, FDEs are an unavoidable link currently; they don't just understand business but also take responsibility for moving data to Agents.
Analyst: Big vendors have more obvious advantages in the data line.
Lai Gangbin: Big vendors certainly have their excellent advantages, such as context from Enterprise WeChat, DingTalk, Feishu, and data from many internal enterprise systems. However, today's ecosystem openness is very good; Enterprise WeChat and Feishu have opened CLIs, supporting comprehensive read/write functions. No enterprise puts all data in one place; openness is prioritized.
Analyst: Are there specific landing scenarios where enterprise customers can clearly see changes?
Lai Gangbin: Take knowledge bases as an example. A typical scenario is forming internal knowledge bases based on large amounts of historical PPTs, Word docs, PDFs, and Excels. Previously, this required complex ETL and data governance; now, handing this to Agents significantly lowers cost and difficulty. Practical results show that governance tasks taking months historically can be compressed to under a week, which is a very visible change for customers. Once the knowledge base is well-built, it helps sales enablement, customer service, and internal training. This is a pragmatic scenario we see frequently this year, ready for implementation.
04、FDEs in the Chinese Market Won't Follow Palantir's Approach
Analyst: Does FDE implementation in China essentially represent an upgrade for traditional ISVs?
Lai Gangbin: Chinese FDEs are indeed different from Palantir. Doing Palantir's approach in China might revert to the previous state of software implementation with 30% gross margin labor outsourcing. At this stage, OPC (One-Person Company) has more opportunities. The linkage between OPC and AI-native organizations is the fastest-moving form currently; many friends are practicing this this year. Previously, their per-capita productivity was hundreds of thousands; with AI empowerment, per-capita income increases significantly, prompting many former corporate insiders to start their own ventures. Customer relationships are existing, and OPC decision-making is quick; a batch of enterprises developed well this year.
Analyst: So FDEs are still the same group of people?
Lai Gangbin: Based on the current status, I think so. The core underlying logic is that entrepreneurs served by FDEs won't progress that fast, and the overall market supply of FDEs is insufficient. Today, as an FDE, you need to understand both AI and customer business. Early FDEs definitely start by understanding one side first. From the business understanding angle, there are many existing talents in the market who hope to enter AI or FDE fields; however, understanding AI and Agents is currently extremely scarce in terms of market supply. People who can truly explain and configure context engineering, skills, connectors, and memory are scarce throughout the market. This is why OPCs starting this year have good per-capita efficiency and economic returns.
Analyst: What kind of person is suitable for FDE?
Lai Gangbin: From the perspective of understanding business and Agents, product managers or product operations personnel are very suitable. With a product background, they naturally have a good understanding of these intersection points.
Analyst: Is it feasible for internal IT staff in enterprises to transition to in-house FDEs?
Lai Gangbin: They must first overcome mental inertia. Traditional software engineering uses deterministic thinking; embracing intelligent agents requires accepting uncertainty, which are two different thinking systems. They must pass this hurdle to truly work on Agents. Those who crossed this barrier in the past year or two are essentially scarce; conversely, people in their twenties without historical baggage adapt faster.
05、After the Agent Platform is Established, New Businesses Will Grow Naturally
Analyst: To what extent has Agent impacted the profit changes in Gemii Technology's private domain operation business?
Lai Gangbin: Customer service is a very obvious point. Previously, it involved massive training, management, quality inspection, rewards, and penalties. Now, the organizational quality of customer service teams is changing: After clients use AI, they should hire fewer frontline executors and more people who understand how to organize customer service knowledge and iterate processes, similar to customer service managers who can daily iterate and self-optimize issues the Agent doesn't understand. This customer service includes not just after-sales, but also precise user marketing, pre-sales consultants, and other broad customer service roles. We used our own intelligent agent platform to build digital employee products, and validation in real customer scenarios showed good results. All those choosing AI upgrades will see significant improvements in profit and efficiency. Of course, organizational upgrades take cycles; when technology isn't the bottleneck, it requires more client acceptance and cognitive evolution.
Analyst: If the enterprise Agent platform can scale agency operation businesses many times over, is this a new direction for the company? For example, acquiring more agency operation teams and handling client businesses with Agents.
Lai Gangbin: We are seriously exploring this; there is definitely an opportunity. Chinese enterprise customers are mostly business-led, with IT providing technological support. Since we can connect products, technology, and business, we have a full chance to replicate this. The core here is that the Agent platform infrastructure must be self-controlled, as all businesses run on it. Gemii's private domain operation business also grows on its own Agent platform, but it's just one vertical direction.
Analyst: How to understand this?
Lai Gangbin: Today we see a bigger opportunity: to create a brand-new foundation for enterprise Agents, whether for private domains or other businesses, everything can run on it. For example, knowledge bases—we had no business in this area last year, but it has developed rapidly this year. It didn't exist in the previous private domain business; once the Agent foundation emerged, new businesses grew naturally.
Analyst: Does building the Agent platform actually not require many people?
Lai Gangbin: Competing with big vendors isn't about resources and headcount, but about scenarios and technical route judgment, focusing on the core team's dedication, and then using AI to amplify hundreds of times.