2026: When AI Agents Started Doing Real Work

A Clear New Signal

July 2026, Shanghai World Expo Exhibition Center. The World AI Conference (WAIC) broke records: over 100,000 square meters of exhibition space, more than 1,100 exhibitors, and over 300 new AI products making their global debut. Tickets sold out within days.

But something was different this year. The buzzwords echoing through the halls were no longer "trillion parameters," "multimodal fusion," or "Scaling Law." Forum after forum, booth after booth, the conversation converged on a single theme: agents.

Zhou Hongyi, founder of 360 Group, put it bluntly: 2026 marks the "Year of the AI Agent." His forecast? AI will fundamentally restructure the world within five to ten years.

This might be the most significant inflection point since the large model frenzy began. AI is moving from "thinking" to "doing."

From Model Race to Engineering Race

For the past two years, the competitive landscape of AI was easy to read: whoever had the bigger model and better benchmarks won—at least until the next launch.

That narrative began to crack in the first half of 2026.

At the AI Engineer World's Fair in San Francisco in early July, one consensus emerged: the competition has shifted from a "Model Race"—chasing ever-larger architectures—to an "Application & Engineering Race"—focusing on real-world deployment and production systems. Nobody cared about parameter counts anymore. The only question that mattered: can this actually work in production?

Around the same time, Perplexity CEO Aravind Srinivas crystallized the shift: "The model itself is no longer the core product. What matters is the 'framework'—the orchestration system that places the model within a powerful ecosystem and matches it with the right tools."

A quiet transfer of power is underway: from model providers to system builders who can weave together models, tools, memory, and workflows.

Agents: From Talk to Action

The core theme of WAIC 2026 was "Intelligent Partners, Creating the Future Together." The organizers chose these words deliberately—AI is no longer a tool or an assistant. It is a partner.

The China Academy of Information and Communications Technology projects that by 2028, over 15% of global work decisions will be made autonomously by AI agents.

The numbers behind this shift are staggering. According to CCTV, China's core AI industry surpassed 1.2 trillion yuan in 2025, with AI penetration in key industries exceeding 80%. Data from Shanghai University of Finance and Economics shows daily token calls hitting 140 trillion in March 2026—up from just 100 billion in January 2024, a thousand-fold increase in two years.

The structure of growth is changing too. Liu Tao, SVP of Kingsoft Cloud, disclosed that his company's employees generate 35 million token calls daily, with "80 to 90 percent" of new code being AI-generated. He noted that the real explosion in inference compute comes from web coding and agent technology—not from training, which dominated in previous years.

Token Economy: The New Unit of AI Value

A deeper shift is taking hold: tokens are evolving from technical metrics into economic units.

China's three major telecom operators are accelerating their pivot from "traffic management" to "token management." China Telecom introduced a Token Security Router for unified multi-model access, quota management, and departmental cost attribution. SenseTime launched "TokenPlan," transforming tokens from a model byproduct into a commercial product. PPIO unveiled an Agent Cloud that cuts token costs by 60 to 80 percent through intelligent model routing.

The battlefield is expanding. It's no longer just about who has the strongest model—it's about who can deliver the lowest token cost. That is a competition governed by real business logic.

Open Source: From Alternative to Main Engine

Another trend accelerated in the first half of 2026: open-source models went from "usable" to "preferred," becoming the default choice for a growing share of enterprises.

Ollama closed a $65 million Series B in July 2026, reaching 8.9 million developers. CEO Jeff Morgan revealed that over 85% of Fortune 500 companies use Ollama's services—including firms in aviation, insurance, and healthcare, sectors that demand strict regulatory compliance.

Benchmark partner Peter Fenton went further: within 18 to 24 months, perhaps even by the end of this year, over 90% of all tokens will be generated by open-source models. His logic is simple: enterprises can now run models that are "good enough" while bypassing the premium fees charged by frontier model companies—and those companies' inference margins will face existential pressure.

Meanwhile, China's open-source ecosystem is rewriting the competitive map. Zhipu's GLM 5.2 is being used by Perplexity to build its new Computer product system. Tencent's Hunyuan Hy3, activating only 21 billion parameters, matches flagship models on agent and office tasks. The catch-up is happening faster than most expected.

Robots Walk Into Factories

If large language models solve the problem of "thinking," embodied AI solves the problem of "doing."

At WAIC 2026, Ant Lingbo unveiled "Full-Stack Brain 2.0," trained on 60,000 hours of high-quality real-world data and already adapted to 20-plus robot configurations from 17 manufacturers. On the exhibition floor, three robots of different shapes and sizes shared a single "universal brain," operating alongside humans in a simulated pharmacy environment—no safety barriers required.

Zhiyuan's humanoid robot Expedition A3 Ultra was named one of the conference's top exhibits. Just before the event, multiple units had already started working in factories. Industry forecasts suggest China's humanoid robot output could exceed 100,000 units in 2026.

But Shen Yujun, Chief Scientist at Ant Lingbo, offered a sobering perspective: "The embodied brain today hasn't even reached GPT-1 yet." This pre-emergence phase is precisely when exploration is most active and industrial deployment most determined.

The Quiet Revolution

Economists often say that people overestimate the short-term impact of technology and underestimate the long-term.

2026 may not produce a single earth-shattering AI event. But what it does produce is more significant than any headline: AI moves from slide-deck "visions of the future" to factory-floor robotic arms, office code generators, and customer-service dialogue engines. It starts doing real work.

Zhou Hongyi's five-to-ten-year timeline might be generous. When token costs drop toward zero, when agents autonomously complete workflows, when robots leave the lab—the much-discussed "GPT moment" may have already arrived, just not evenly distributed.

And 2026 is when that distribution begins to accelerate.