The agent gold rush flipped in one summer
Last year, AI agents were the hottest ticket in venture capital. DeepSeek and Manus turned “the year of the agent” into a funding frenzy — projects with no product and no revenue raised rounds, and some investors had to commit to two follow-on rounds just to get a seat at the table. By the first half of 2026 the sector still looked unstoppable: 83 agent deals in China’s private market, up 131% year over year.
Then summer arrived, and the mood froze. “We are not looking at agents anymore.” “AI application projects are stuck internally.” Those lines spread from a few cautious firms to a collective sentiment across the entire venture ecosystem. In a matter of weeks, most VCs redirected attention to embodied AI and AI-for-science. Agent founders — regardless of pedigree — now face the same grilling: “How many users? What is your projected paid conversion? What do you do if a model lab ships your feature?”
This is not a China-only story. It is the loudest early warning that the agent layer is being commoditized faster than anyone expected — and it contains a playbook for founders everywhere.
Why the bubble popped: three structural forces
The freeze was not a sentiment shift. It was three hard realities colliding.
1. Model capability ate the middle layer. Yupp.ai, a crowdsourced model-evaluation platform with $30M+ raised and 1.3M users, shut down in April. Its core value — humans rating chat models — became obsolete as models got too good to need it. When the underlying capability moves that fast, any business built purely on top of the model’s old limits evaporates.
2. General-purpose agents became a big-tech game. Consumer agent traffic is sliding: Manus fell from 28M to 23M monthly visits, Genspark from 15M to 11M, and smaller players like Dokie and Flowith dropped 20-30%. Meanwhile Tencent’s WorkBuddy hit 20.97M monthly PC visits in June and ByteDance’s Trae IDE reached 12.79M. The analyst logic is brutal: consumer agents enjoy no loyalty — the moment a smarter, cheaper model arrives, users switch.
3. Agent SaaS broke the SaaS math. Traditional software scales with near-zero marginal cost; gross margins run 75-85%. An agent, by contrast, burns tokens and inference compute on every single task. More users means more token spend, not more margin. That is why successful agent companies stopped charging per token and started bundling: Beisen charges “points,” Yunfu bills by call duration, and LibTV meters membership credits. The old 70% gross-margin slide deck no longer survives contact with reality.
The escape routes: vertical B2B and going global
The survivors are not waiting for the hype to return. Two paths are actually producing cash flow.
Vertical B2B with real delivery. One funded startup automates customs declaration forms — one of the most labor-dense, repetitive jobs in trade. A human fills roughly 10 forms a day at a cost of about 1,000 RMB each; the agent handles 100+ a day at 600-700 RMB. That is the pitch that lands: “use us and hire fewer people.”
Delivery now splits into two modes, mirroring old-school software services. Standard SaaS works where data and workflows are uniform; but enterprise clients with strict data-security requirements will not send sensitive data to a cloud platform. That gap created the FDE (Field Delivery Engineer) model — teams embedded on-site to retrofit an agent around a customer’s existing processes, in projects that run months or even a year. Yaowu Technology, a manufacturing-focused agent firm, cut over 90% of one client’s pre-sales customer-service headcount with 97% response accuracy. Their take: the real moat is not the model, it is trust and organizational fit — clients hand over proprietary data only when they intend to stay.
Overseas, where paying customers actually pay. Dify, WorkMagic, and Ace Studio have quietly crossed $10M ARR by going after global markets from day one. One founder sells a “Jarvis for kids” — a wearable audio egg that transcribes classroom audio on-device, then sends it to a cloud agent that generates personalized tutoring. In China, parents were burned out by marketing and would not convert; at a Singapore expo, families bought on the spot. Southeast Asia feels like China 20 years ago: fewer than 20 competitors per niche (versus thousands at home), more pricing power, and customers who believe what they see. The cost is brutal localization — different school systems (PSLE versus SPM), different ethnic and religious composition per market, WeChat versus WhatsApp — plus local incorporation for trust.
What this means for the global agent market
Read the China signal, because it is usually an early warning for the rest of the world:
- The “agent as API wrapper” thesis is dead. Whatever the frontier model can do tomorrow at one-tenth the price, your thin wrapper cannot defend.
- Margins, not hype, decide who survives. If your unit economics depend on per-token markup, a cheaper model can kill you overnight.
- Moats migrate to what models cannot copy: proprietary industry data, localized context, delivery capability, and long-term trust.
- Big labs will eat horizontal apps. The winning startups are those embedded deep enough in a specific vertical or region that the model lab cannot justify going after them.
The agent winter is not the end of the category — it is the end of the free ride. Related reads: how agents infect each other and why reasoning moats are collapsing.
What founders should do this quarter
- Pick a vertical where the data is messy and the model cannot see it. Customs forms, manufacturing pre-sales, and school tutoring all won because the industry context lives outside the LLM’s training data.
- Fix unit economics around outcome, not tokens. Bundle pricing, outcome-based contracts, and delivery wrap the token cost into something clients can budget for.
- Build a delivery motion, not just a product. The companies that survived are the ones willing to sit inside a factory or a customs office until it works.
- If you are a consumer team, go global early. The market with 20 competitors beats the market with 2,000 — even if you must register a local entity to earn trust.
- Assume your current model advantage has a six-month half-life. Everything you build must compound an asset the next, cheaper model cannot erase.
The agents that survive are not clever wrappers. They are companies that dug deep enough into a messy, human, high-trust problem that the model frontier passed them by instead of swallowing them.
