From Open Source to Agents, from Organizations to Individuals: Notes from the AIEC Conference Floor
Summary
By mid-2026, model and application vendors had reached broad consensus around Claude Code-style Agents, with the main battleground expanding from model leaderboards to executable workbenches. The line between Code Agents and general-purpose Agents is disappearing. Nearly every product now looks like “a project pane on the left, a chat pane in the middle, and an execution pane on the right,” with MCP and Skills layered on top; Harness and Loop represent the corresponding execution and training logic. Once the interfaces converge, the differentiators become context, cost, usability and each company’s installed ecosystem.
What enterprise AI actually lacks is usable context and organizational readiness—not executive attention. Midea’s observation is that the digital era was characterized by urgency below and indifference at the top, while the AI era has become “the boss is relatively eager, but the people doing the work may not be.” Without data, permissions, processes and KPI redesign, models have “nothing to latch onto”; the digitalization homework enterprises skipped now has to be done all over again.
Deployment will not follow a single “move fast in small steps” playbook: mature core businesses may be driven down from top-level metrics, while innovation projects bubble up from the edges. Midea argues that AI’s value no longer needs to be litigated; with budget and cultural support, companies can go straight at the core and replace legacy processes with advanced capabilities. Tencent WorkBuddy, meanwhile, evolved from an individual side project into an internal breakout product, showing why companies need to accommodate strategic execution and grassroots experimentation at the same time.
Jensen Huang’s $20T application layer translates to roughly RMB140T, but market size does not mean model companies can own the entire stack. Once a vertical produces a breakout company quickly enough, the model vendor may move in. Startups have the inverse advantage: a 20–30-person team may invest far more deeply than the 2 people a large company has assigned internally. 庄明浩 is more excited by “100 million small companies, not 200,000 $100M companies.”
Multi-Agent systems are moving from a product feature to a 3rd Scaling path, with a division-of-labor logic that increasingly resembles human organizations. Beyond token efficiency and long context, Kimi has added multi-Agent clusters in which different roles carry different contexts and capabilities, with reviewers closing the loop. Enterprises likewise should not route every task to the most expensive model; they need a routing layer that balances outcomes and cost. A “nation of geniuses in a data center” still needs organizers.
The first enterprise value of open source is business continuity and data sovereignty; monetization will shift toward a bundled model. Traditional factories cannot tolerate the power switch for a critical model being held by someone else, nor can they hand the equivalent of the Coca-Cola formula—their context—to the entire world. Open-source models can bind infrastructure below and Agents above, then monetize through tokens, plans and tools, because a durable business is difficult to build on goodwill alone.
Model competition has entered a monthly-scale iteration cycle of roughly 1.5–2 months, making the “keep development secret and unveil the big move at year-end” model increasingly dangerous. The discussion cited successive releases including Kimi 3.2 and MiniMax M3, summarizing the pace as: “Hold it down for 5 seconds and the world has moved 5 years.” That is why even well-funded companies with ample talent and compute can still lose momentum through closed development and long feedback cycles.
As software becomes increasingly generative in real time, the scarce assets will sit at 2 ends: hardware constrained by physics, and human judgment that cannot be outsourced. Power, chips and equipment have little short-term supply elasticity and may be re-rated. At the other end are emotional reactions, taste, trust, presence and the marks of making something by hand. Even if AI writes “the most surprising line in the room” for you, you still need to keep your first judgment: “If you care about it, it has meaning.”
Deep dive
1. AIEC Channels Scattered Hot Spots into 3 Themes: Open Source, Agents and Organizations
庄明浩 said AIEC 2026 had deliberately assembled industry analysts, academics, frontier model teams and traditional enterprises, aiming to cover open source, Agents, organizations and industrial transformation rather than just model scores.
His broad conclusion was that “there seems to be no standard answer at this point.” Every company is advocating its own path; as with training a model, “you train a little and have no idea what it will become,” so the only option is to build and watch.
高飞 noted that terms such as world model and Physical AI were appearing on trending lists at random. The 2 then discussed how Agents, content creation, model companies and a potential SpaceX listing might be linked by investment narratives and the secondary markets in China and the US.
2. Technical Invention Is Only the Starting Point; Organizational Fit Determines the Speed of Productivity Gains
庄明浩 used the steam engine as a reference point: 30–40 years may separate a technical invention from a visible increase in GDP. The real delay lies in adapting organizations, production relationships, transaction structures and the surrounding industrial chain.
AIEC’s 3 examples covered different scales: Midea, with its complex global operations; Inspur, embedded in the IT infrastructure chain; and a team using OPC to transform traditional energy companies. Even a 2%–3% efficiency gain has tangible value.
The cases point to a common conclusion: enterprise AI is no longer merely a technical solution or procurement decision. It is also a question of culture, processes, permissions and how people reconnect with existing institutional arrangements.
3. The First Enterprise AI Bottleneck Is Context Engineering, Not Model Procurement
高飞 relayed the change described by Midea’s 魏总: digital transformation had long been seen as “valuable, but impossible to measure immediately,” leaving C-level executives verbally supportive but not necessarily willing to act. After sustained exposure to large language models, the question has shifted from “should we do it?” to “how do we do it?”
魏总 summarized the deployment challenge as context engineering: without completed digitalization and structured data, the model has nothing to work with. The AI boom is forcing companies to complete the internal work they had postponed.
庄明浩 said China and the US face the same friction. Even when models are ready, enterprise data, processes, architecture, employee habits and KPIs are not. FDEs and similar roles are therefore needed to connect the model’s capability boundary with real-world deployment.
4. Digital Employees Force Companies to Rebuild Permissions, Data Fences and Cost Accounting
Inspur’s basic logic is that once an Agent becomes a “digital employee,” it also needs permissions, data fences and a dedicated environment for its runtime data. As Agent concurrency rises, multi-Agent management, hardware-software coordination and enterprise data isolation all become infrastructure problems.
Employees’ existing work must also be organized into context that AI can understand and use. Inspur is using internal hackathons, a Skill Hub and its IT team to accumulate capabilities, then plans to reuse them with customers that buy its hardware and infrastructure.
The real difficulty is that the new system is not naturally compatible with legacy KPIs. Different functions also vary widely in their willingness to adopt AI and in the returns they can generate from redesign.
Cost constraints have replaced simple “maximize token usage” thinking. Previously aggressive companies in China and the US are now recalculating token efficiency—how many fewer tokens are needed for the same task, or how much more complex a task can be solved with the same token budget.
5. Core Business Moves Top-Down; Edge Innovation Moves Bottom-Up
Midea has pushed back on the traditional “move fast in small steps” approach: digitalization had to prove value at the edges, but AI’s usefulness is no longer in dispute. If leadership, budget and culture are aligned, companies can start with core systems and use advanced capabilities to replace legacy processes.
Tencent WorkBuddy followed the opposite path. Inspired initially by the emergence of Claude Code, it began as a personal side project at the edge, then became an internal breakout product after performing well in small-scale trials—an example of AI-native innovation moving upward from below.
Drawing on his own experience at internet companies, 庄明浩 distinguishes between 2 types of work. The established social business gets top-down direction around retention and conversion; incubation projects use accumulated technology to search for “emergence,” launching small products first and watching user feedback, peer reactions and even whether anyone copies them.
The 2 paths can run simultaneously. “No standard answer” does not mean stopping; it means allowing core businesses to operate under strategic constraints while edge projects retain room for low-cost experimentation.
6. Claude Code-Style Workbenches Have Become the Industry’s Common Language
庄明浩 had expected model companies and Agent teams to tell 2 different stories. After hearing them, he found that—with the exception of Kimi’s particular emphasis—the presentations pointed to the same consensus: the form represented by Claude Code is unifying model and product roadmaps.
The form is not merely an interface; it also feeds back into model training. Harness and Loop follow similar organizational and training logic, while the boundary between Code Agents and general-purpose Agents becomes increasingly blurred.
Product names and structures have converged as well: Claude Code/Cowork, Kimi Code/Kimi Work, Trae Solo/Trae Work, Tencent CodeBuddy/WorkBuddy, and Alibaba Qoder/Qoder Work all use some version of a project pane, chat pane and execution area.
This creates a new battleground beyond chatbots. Tencent, ByteDance, Alibaba, Kimi, MiniMax, 智谱, OpenAI, Anthropic and Google are all putting marketing, compute, pricing and distribution resources behind the category, and the competition is far from over.
7. As Interfaces Converge, Ecosystem Compatibility and Low-Friction Delivery Become More Valuable
庄明浩 compared the trend to the iPhone standardizing the shape of the smartphone. Agents may standardize software into similar workbenches; once the market reaches the stage where “no one can accuse anyone else of copying,” buyers will naturally look for the differences that actually matter.
The most direct point of differentiation is each company’s installed ecosystem: Tencent connects to Tencent Docs and Tencent Cloud; Alibaba connects to Alibaba Cloud, e-commerce, 千问 and 灵光; ByteDance can link Volcano Engine, multimodality and Douyin.
庄明浩 used a bridal photography studio in Changchun as an example. A friend wanted to build a marketing Agent, knew only Doubao and Yuanbao, and had no engineering background. The recommendation should be based not on model rankings but on usability, compatibility, security, trust and fit with the specific workflow.
8. Most Users Want a “Phillips-Head Screwdriver,” Not Infinite Models
The bridal photography example made 高飞 realize that many users do not need a broad capability set; they need a defined workflow. A model combined with Cowork and Skills may complete the task, but handing that assembly method directly to a nontechnical user still leaves them unable to use it.
The shared value of 背背佳, electronic dictionaries and learning machines is that they package functions users could assemble themselves into a defined answer. People pay not because the underlying capability is irreplaceable, but because the product works out of the box.
The key metaphor from the discussion was that today’s AI “is neither a screwdriver nor a pair of chopsticks.” Coding may be an exception. Infinite model capability is an advantage, but an ordinary user may simply need one clearly defined Phillips-head screwdriver.
9. The $20T Application Layer Will Not Automatically Accrue to Model Companies
庄明浩 cited Jensen Huang’s “5-layer cake”: the application layer has an upper bound of $20T, or roughly RMB140T—about the size of China’s GDP. The unresolved question is whether model companies will take the entire layer, or whether it will produce large numbers of super-applications and small companies.
The discussion used Cursor as an example. It reportedly accounted for roughly 50% of Anthropic’s revenue at one point in its early days, before Anthropic used Claude Code to “kill its own customer.” Cursor was reportedly acquired by xAI, may still be growing, and is also trying to build its own model on top of open-source models.
The legal industry is following the same pattern. Harvey and another legal AI company whose name was unclear in the transcript quickly reached $100M–$200M in scale, after which Claude launched a product for legal use cases. Once a vertical is large enough, the foundation model company has an incentive to move up the stack.
Startups benefit from an “inverse mismatch”: a team may be able to put 20–30 people into a narrow vertical, while a large company has only 2 people assigned internally. The real barriers to total domination by giants are organizational entropy, capability boundaries and whether models can generate personalized products on their own.
10. The Cement Agent Shows That Vertical Moats Lie in Data and On-Site Deployment
A cement Agent does not rely solely on a large language model. It looks more like a specialized time-series model for materials science. As long as language remains the primary training corpus for large models, it will remain naturally separate from industrial process data.
Even if foundation models eventually absorb that data, the extent of their generalization remains uncertain. The more immediate moat is that a model company may not send an on-site deployment engineer into a single cement plant for 3 months.
That leaves a sharp unresolved debate over the next step: continue scaling language models, move toward world models or AI for Science, pursue self-evolution, or use multi-Agent systems to solve complex tasks. No path has won.
11. Multi-Agent Systems Are Both a Scaling Path and a New Organizational Structure
高飞 divided the future into 2 paths: centralized intelligence that continues to self-evolve, and systems in which no single model needs to do everything because multiple Agents collaborate on complex objectives.
高飞 traced Kimi’s trajectory: at the beginning of the year, the focus was on 2 forms of Scaling—token efficiency and long context. The 3rd is now multi-Agent clusters, aligned with the capabilities Kimi has emphasized from 2.0 and 3.0 through 3.2.
Multi-Agent systems give different roles different contexts, capabilities and responsibilities, then use an integrator and reviewer to close out the work. This is converging with both the way Harness and Loop train models and the way humans design workflows.
Enterprises need a routing layer for model calls as well, rather than sending every task to the most expensive model. The “nation of geniuses in a data center” described by Anthropic’s founder still faces a Manhattan Project-style problem: geniuses need division of labor and organization.
12. A “Nation of Geniuses” Still Needs Engineers, Coordinators and Organizational Boundaries
庄明浩 cited hiring statistics from an AI lab: most roles are not pure research positions, but infrastructure and back-end engineers from systems such as AWS and Meta. “Geniuses can’t even solve the problem of feeding themselves, let alone the problem of feeding GPUs.”
The discussion also covered Meta monitoring employees’ mouse and keyboard activity, laying off 8,000 people, and Alexandr Wang pulling several thousand engineers into an AI lab for task labeling. Employee resistance was summarized as feeling “like being thrown into a concentration camp.”
庄明浩’s response was that this approach can optimize only known problems. Turning existing employees into Skills may lower costs, but it cannot automatically solve unknown problems; Meta’s existing social and glasses businesses are not simply efficiency shortfalls.
Wall Street buys “the future” and may not reward straightforward quality or efficiency gains. 庄明浩 is therefore less optimistic about the trend of continuously buying GPUs, while 高飞 noted that getting the AI business, cloud and legacy franchise to win market recognition simultaneously is itself an extremely difficult task for incumbents.
13. Monthly Iteration Is Bringing the “Closed-Door Big Reveal” Era to an End
Technology giants now span energy, data centers, chips, models, cloud, B2B and B2C, while a competitor at one layer may be a partner at another. 庄明浩 called the landscape “a mess in a pot,” with everyone refusing to turn back—“not before hitting the wall, and not even after hitting the wall.”
The discussion cited Musk-related businesses, Nebius becoming Anthropic’s cloud inference provider and plans for compute in space. Industry boundaries have yet to harden; everyone wants to hit the wall once before confirming their place in the ecosystem.
庄明浩 joked that the companies most likely to train models badly may be those that were “too good at keeping their work secret,” citing xAI and Apple. They had no shortage of money, people or long-term commitment, but during 3 months of closed development, external technology may have gone through several iterations.
The current update cycle is roughly 1.5–2 months. The discussion listed Kimi 3.2, Claude 5.2 and MiniMax M3. SSI has been closed-door for 2 years; unless the paradigm changes so radically that it is no longer based on the Transformer, the “big reveal” model is difficult to sustain.
14. Open and Closed Source Are Moving Back from Extremes Toward a Dynamic Middle Ground
高飞 sees DeepSeek as the 1st wave: low-cost inference opened the market and also gave rise to the all-in-one appliance narrative that later fizzled. The 2nd wave was Qwen, with many overseas models fine-tuned on it; more recently, teams have also been fine-tuning models based on Qwen 3.5.
It is now difficult to identify a new standard-bearer. The early narrative held that training was expensive and scientists were scarce, making closed source the inevitable winner; last year the pendulum swung to the opposite extreme represented by Chinese open-source models. This year, both sides are searching for a middle ground.
Closed-source companies are learning from open ecosystems and releasing non-core models. Open-source vendors must decide whether to keep their largest and most Pro versions open. Kimi, MiniMax and 智谱 remain in competition, and the balance will continue to swing.
The discussion also used the US government’s ban on Claude’s latest 5th-generation models as an example of the side effects of one-size-fits-all regulation. Neither regulators nor vendors have a simple answer for how open systems should be, or which capabilities must remain controlled.
15. For Traditional Enterprises, Open Source First Means Staying Online and Keeping Data Private
高飞 compared the issue with factory power supply. Once AI is embedded in a core production line, a company cannot tolerate the power switch being held by someone else. If manufacturing stops, the plant may not simply restart later; “one batch of material could all be scrapped,” making an in-house “generator” essential.
Enterprise context is a moat as well. Handing all of a company’s corpus and data to an external model is like giving the Coca-Cola formula to the entire world. Even if outsiders consider it less mysterious, the company will not volunteer it.
Open source therefore does not mean acting as a “bodhisattva.” It supports a bundled business model: sell tokens, token plans and tools; connect downward into infrastructure and upward into Agents. Open ecosystems can build influence, but long-term commercial viability cannot depend on goodwill alone.
16. As AI Moves Deeper into White-Collar Work, Human Reaction, Taste and Presence Become More Valuable
The happiness psychologist 彭院长 emphasized kindness, experience, presence and individuality. 庄明浩 noted that even highly educated parents around 李飞飞 are asking what their children should study, showing that career anxiety no longer has a clear intellectual safe haven.
庄明浩 placed Go, simultaneous interpretation, Coding, image creation and video creation on the same trajectory: once technology crosses the threshold, it becomes unstoppable. The helplessness produced by Claude 5 is that, in some white-collar tasks, a person’s best contribution may be to stand aside and watch rather than interfere blindly.
The 5 games between 李世石 and AlphaGo were used to describe the psychological process: first doubt, then worry, and by game 3, “the Dao in one’s heart is shattered.” This generation may be like textile workers in the steam-engine era, witnessing productivity rise while also seeing expertise, careers and stable lifelong relationships pulled apart.
People can still retain their first emotional reaction, taste and trust. AI may rush to write “the most surprising line in the room,” but likes and dislikes cannot be outsourced. Hand-built PowerPoints, live meetings, long-form podcasts and expression untouched by pre-processing are valuable precisely because “if you care about it, it has meaning.”
17. The Ecosystem Will Ultimately Stratify, and Physically Constrained Hardware May Be Re-Rated First
The discussion was willing to forecast only the next 3–6 months. One set of materials had to be rewritten continuously from January through March, late April and mid-June; information was overflowing every 1.5 months, making confident 1–3-year calls unrealistic.
庄明浩 hopes the industry will grow from “a single strange plant” into a real forest. After repeated rounds of competition, models, Agents, infrastructure, vertical applications and enterprise services should develop clearer roles and boundaries.
The closing hardware view was that “software is being produced in real time,” while hardware remains constrained by manufacturing cycles and physical conditions. It depends on power below and connects supply with demand above; with an inelastic supply curve, “old-economy stocks” and hardware hubs may receive another round of re-rating.