---
title: "WAIC Opens Today: We Deconstructed the Full Path of the Agent Economy Two Days Early"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-07-17T02:57:05+00:00"
canonical: "https://ffcap.cn/en/research/src-20260717-01html"
source: "https://uniqueresearch.substack.com/p/src-20260717-01html"
language: "en"
---

# WAIC Opens Today: We Deconstructed the Full Path of the Agent Economy Two Days Early

[![](https://substackcdn.com/image/fetch/$s_!nhYg!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b9d919b-452d-49ab-aac1-94672f652c75_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!nhYg!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1b9d919b-452d-49ab-aac1-94672f652c75_2048x1152.png)

_Original · Unique Research / 非凡产研 · 2026-07-17 · Chinese source: https://view.inews.qq.com/a/20260717A043MO00_

_Editor's note: This is a complete English rendition of the source article. Speaker attributions and quoted predictions are retained as the speakers' own statements. Source images are not processed per task scope._

\---

**Unique Research Awards**

"AI is no longer satisfied with being a 'tool.' It is competing for a 'job.'"

The highly anticipated WAIC 2026 officially opened today. Ahead of this major industry gathering, the "2026 Unique Research Awards · Shanghai · AI Business Conference," hosted by Unique Research, had already concluded successfully. The first day of the summit focused on common industry challenges, examining practical paths for AI implementation and connecting business workflows into closed loops; the second day elevated the lens to the full Agent economy value chain, covering key topics such as business models, infrastructure, vertical use cases, and overseas expansion. The 7th Unique Research Awards were presented on-site, and an afternoon track on AI imaging commercialization was also held. A clear consensus emerged throughout the summit: AI's value is no longer limited to single-point tool capabilities, but is gradually becoming an independent role collaborating within enterprise organizations.

**Speaker | Wu Wei (吴畏), Founder, Unique Research**

Wu Wei opened with a blunt judgment: the AI industry is collectively crossing an inflection point, moving from "you ask, I answer" Copilots to Long-running Agents that "finish the job." This is not any single company's storytelling—earlier this year, Alibaba, Tencent, Moonshot AI, and ByteDance all upgraded their coding Copilots to Work Agents in nearly the same quarter. When trajectories converge this neatly, it means the curve itself has pushed us to this point.

The data is equally direct: non-developer users on Codex grew more than 5x in six months, and the share of tasks lasting 8+ hours surged from 2.1% to 25.6%—nearly a tenfold increase. Anthropic's own comparison is even more striking: for the same article, Chat mode requires 13 rounds of back-and-forth, while Agent mode finalizes it in one sentence. Industry research corroborates this: over half of the participating products already have three or more long-horizon core capabilities. "Digital employees" have become an independent category. Nearly 20% of products are starting to charge for "virtual labor" rather than by API call volume, and one-third of traffic flows through desktop clients, group bots, and other channels that don't require human supervision. In plain terms: "give it a goal and let it run to completion" is already a product form that people are genuinely paying for—not a slide deck.

But Wu Wei didn't stop at "this is impressive." He raised the harder question: how do you take an Agent from Demo to a Product that stays online and has someone accountable when things go wrong? His answer was a four-capability framework: continuous context understanding, evidence-based judgment, genuine execution, and always-on controllability. This solves not an IQ problem, but a trust problem. The management logic must also change—from "managing people" to "managing context." The memory system is not a chat-record warehouse; it must structure execution experience into a four-layer pyramid—context, conversation, project, skill—consolidating into "muscle memory." Trust is designed, not given: start with tight permissions, loosen them as it proves itself repeatedly correct, and immediately revoke after a single error.

"For a 100-step task, even if single-step accuracy reaches 99%, overall success rate drops to just 37%. That's why Multi-Agent is not a panacea—it's only worth it in scenarios where genuine role conflict and adversarial dynamics exist. Most enterprises haven't even gotten a single Agent running steadily; writing the job description clearly is the real priority."

He also pointed out a severely underrated battleground: Agent Runtime. A demo can be built in an afternoon, but a production product needs to integrate models, sandboxes, browsers, search, logging, state, and permissions—seven vendors, each individually inexpensive, but the combined technical debt is crushing. Just as the website era needed Cloudflare and the payments era needed Stripe, this infrastructure layer for the Agent era is still empty; whoever fills it first wins. Market data also signals: digital employee ACV ranges from 100,000 to 5 million RMB, renewal rates exceed 90%, and the business model has shifted from "selling software" to "selling outcomes." He judges that three inflection points may collide within the next 12 months—models absorbing today's Harness scaffolding, 1.4 billion monthly active consumers developing the habit of background autonomous execution, and Agents beginning to directly negotiate transactions with each other. The ultimate organizational form is "a few people managing a large cohort of digital employees." But he emphasized: Agents only take work that can be written as formulas; decisions requiring intuition, judgment, and accountability are, instead, more rightfully left to humans. The winner is not the one with the strongest model, but the one who truly embeds intelligence into the business timeline and prepares the organization for this division of labor upgrade.

**Speakers | Wu Xiankun (吴显昆), Co-founder & CEO, Kuse AI; Xu Leyang (许乐洋), Co-founder, Seekee AI; Cameron Wang, GM, Ecosystem Partnership Center, SenseTime; Ian Li (李忆), Business Manager, Greater China New Customer Business, Google; Gu Xuebin (顾学斌), Founder & CEO, WeMeet**

**Host | Zhuang Minghao (庄明浩), VP & Chief Strategy Officer, Quto Technology**

The most cutting line from this panel was: the biggest problem with AI products right now isn't that nobody uses them—it's that you can use them and still can't collect revenue. Consumer traffic looks large, but token costs scale linearly with user volume, eating gross margins alive; a swarm of freeloading users click a few times and leave, and the bill left behind exceeds revenue. B2B ACV can go higher, but every enterprise's processes are "non-standard parts" with no unified API; internal business documents aren't even organized. Building Agent deployment isn't installing software—it requires first doing a round of management consulting for the customer. The deeper problem is that customers are increasingly unwilling to pay for "how many times you called the model." What they want is "what did you solve for me?" This means the entire industry must shift from selling tools to selling outcomes—but how to define outcomes and quantify ROI, no one in the room had a standard answer for.

Going overseas is seen by many as a lifeline. Emerging markets like Latin America and Southeast Asia are indeed willing to pay for vertical-scenario AI products. But the truth is that the window of "any AI gadget will monetize" overseas has long passed. What's emerging now are vertical Agents embedded deep into specific workflows—helping creators grow followers, accelerating gene sequencing, building 3D models for pets. Only by embedding into someone's production process can you escape homogenization. And the pitfalls of going overseas are deeper than they appear: compliance filings, cross-timezone collaboration, local freelance management—these hidden costs eat profits by the minute. On acquisition, B2C is cheap but converts poorly; B2B has few keywords but high ACV. Choosing a channel is essentially choosing "whose money you want to earn."

To truly commercialize, technically you need model orchestration—routing tasks to models at different cost-performance ratios, using older models' price drops to control costs. On delivery, private deployment plus industry templating is the major trend; enterprises want both data security and out-of-the-box use. But harder than technology is organizational change. An Agent isn't an IT project—it's a management revolution. If business leaders don't deeply participate and there's no dedicated operations team, it becomes a zombie system in three months. The entire discussion landed on one validation proposition: can AI raise an enterprise's profit margin from 20% to 30%? If you can quantify this, achieve it, and collect the revenue, the business model is truly established. Otherwise, it's all talking to yourself.

**Speakers | Han Yunyun (韩云芸), COO, EverMind; Si Hongxing (司红星), Founder & Chairman, WanJing Security; Shen Junxiao (沈俊潇), CEO, Memories.ai; Zhang Minsong (张民松), Distinguished Researcher & AI Lead, Shandong University & Century Tiantian**

**Host | Zhao Liang (Abner) (赵亮), Partner, Unique Capital**

The disagreements on this panel were more interesting than the consensus. What exactly is Memory? Some say it's a dimension of agent subjectivity, complementary to reasoning logic, and should build a self-evolving "personality model." Others, from a practical standpoint, slice memory into three layers: in-session cache, mid-task tracking, and team long-term assets. Education scenarios emphasize that memory must embed disciplinary boundaries and compliance constraints, never crossing lines. Visual memory is treated as infrastructure for physical AI, because video isn't a native input for large models and requires dedicated encoding, compression, and indexing. Different formulations, but they're asking the same question: model contexts are getting longer—should Memory get thinner or thicker? The answer is counterintuitive: text memory can thin, but scenario-based structured memory must thicken.

On the technical roadmap, the industry is moving from single RAG or pure Markdown files to hybrid architectures—RAG for recall efficiency, structured files for precision. Visual memory teams have already compressed encoding costs for thousands of hours of video to under 100 RMB, with local deployment available; the threshold for AI understanding the physical world through cameras is dropping off a cliff. Interestingly, in resource-constrained scenarios—offline environments that can only run 7B small models—long-chain tasks easily cause the model to "spin," repeatedly calling tools without breaking out. One team has designed an "external observer" to detect such loops, automatically distilling structured memory entries, then using a multi-dimensional scoring model to dynamically judge which memories are valuable. These innovations all say one thing: Memory architecture must be tightly coupled with task complexity, model capability boundaries, and deployment environment constraints. There is no one-size-fits-all solution.

The real difficulty isn't technical—it's "boundaries" and "trustworthiness." In heavily regulated domains like education and cybersecurity, Memory isn't just about "remembering more"; it's about "knowing what not to remember and what not to say." Syllabus stage boundaries, red lines on territorial issues, security decisions in attack attribution—these are hard constraints. What customers want isn't a smarter model, but a memory system that continuously accumulates experience while being auditable and controllable. So the industry no longer argues about "whether agents need Memory"; they argue about who can weld general-purpose memory infrastructure tighter to vertical business logic, who can manage memory more efficiently under extreme conditions like small models, offline, and high security requirements. Those who emerge will necessarily be "dual-domain" players who understand both memory technology and industry context; single-dimension players will quickly be left behind.

**Speakers | Kong Weigang (孔维刚), Founder & CEO, Fenghuolun Yingtu; Duan Ran (段然), CEO, Xingqiong Fangzhou; Hu Xiaoping (胡晓平), Group VP, Flexiv; Jia Xindong (Slade) (贾昕东), Overseas 2C Market Lead, Shanghai Whale AI Robotics; Li Yang (李杨), Partner & VP, Shenwu Technology**

**Hosts | Kang Zhengzhong (康正中), Shenzhen Lead, Unique Research; Huang Jingrui (Jerry) (黄璟睿), VP, Unique Capital; Wang Chaochao (CC) (王朝超), Partner, Unique Capital; Jiang Zhiqiang (江志强), Venture Partner, Helian Capital AI Fund; Wu Wei (吴畏), Founder & CEO, Unique Research; Huang Shufei (Sophie) (黄姝菲), Co-founder, PraxisGrowth; Zhang Zifeng (ARK) (张子峰), Founder & CEO, iSpiral Inc.**

---

Original publication: https://uniqueresearch.substack.com/p/src-20260717-01html
On-site reading page: https://ffcap.cn/en/research/src-20260717-01html