There is a recurring pattern in technology where the industry collectively fixates on one metric, one benchmark, one number that everyone agrees determines who is winning. In the early days of the internet, it was page views. In the smartphone era, it was app downloads. In the first phase of generative AI, it was parameters — the sheer size of the model, measured in billions, then trillions, as if intelligence were a function of scale alone.
At Baidu Create 2026, held this week in Beijing, CEO Robin Li stood on stage and effectively declared that era over. The model wars, he argued, have reached a point of diminishing returns. The frontier models from OpenAI, Google, Anthropic, DeepSeek, and Baidu itself are now so capable that the competitive advantage no longer lies in building a slightly larger or slightly smarter model. It lies in what you do with the models you already have. Specifically, it lies in agents — autonomous AI systems that can plan, act, and complete complex tasks on behalf of users.
This is not a niche technical distinction. It is a fundamental reframing of what AI means for society, for business, and for the billions of people who will interact with these systems in the coming decade. And it is coming from a company that, for all its struggles in the Western consciousness, remains one of the most sophisticated AI organizations on the planet.
The End of the Parameter Race
Baidu's argument is grounded in observable reality. The gap between the best models and the very best models has narrowed dramatically over the past year. A user interacting with GPT-5.4, Claude Opus 4.6, Gemini 3.0, or Baidu's own Ernie 5.0 would be hard-pressed to identify which model is generating the responses in a blind test. They are all capable of sophisticated reasoning, complex coding, nuanced analysis, and creative generation. They all pass the bar exams, ace the medical boards, and write poetry that makes English professors weep.
What they cannot do, reliably and at scale, is complete multi-step tasks in the real world. Ask any of these models to plan a complex business trip, negotiate a vendor contract, or manage a content calendar for a month, and you will get a beautiful plan that sits in a chat window and does nothing. The model thinks. It does not act. And in a world where execution matters more than ideation, that is a profound limitation.
Li's thesis is that the next phase of AI value creation will come from bridging this gap — from turning models that think into agents that do. And the companies that master this transition will define the next decade of technology, just as the companies that mastered the mobile transition defined the previous one.
What Baidu Is Actually Building
Baidu's agent strategy is not theoretical. The company demonstrated several concrete products at Create 2026 that illustrate what agentic AI looks like in practice, and the vision is both ambitious and grounded in China's particular technological and regulatory environment.
The flagship demonstration was a personal assistant agent that operates across Baidu's ecosystem — search, maps, cloud storage, smart home devices, and third-party services integrated through Baidu's platform. Unlike a traditional voice assistant that responds to discrete commands, this agent maintains persistent context, learns user preferences over time, and proactively initiates actions based on predicted needs.
The example that resonated most with the audience: the agent notices that you have a meeting across town in ninety minutes, checks real-time traffic data, realizes you will be late if you leave at your usual time, and automatically sends a message to the meeting organizer suggesting a ten-minute delay. It then books a rideshare, adjusts your smart home to prepare the house for your return, and queues up the documents you will need for the meeting based on your calendar and recent activity. All of this happens without a specific command. The agent is operating on your behalf, using your preferences as its guide.
This is not science fiction. Baidu has been running a closed beta of this system with hundreds of thousands of users in China for several months, and the company reported that users who adopt the agent complete an average of forty percent more tasks per day than those who do not. The metric is deliberately vague — what counts as a "task" is not defined — but the directional signal is clear. When AI stops waiting for instructions and starts anticipating needs, productivity changes in ways that are difficult to reverse.
The Rise of the Super Individual
Perhaps the most striking theme of Li's keynote was his repeated use of the phrase "super individual." In Baidu's vision, AI agents do not replace humans. They amplify them. A single person with a well-trained agent, Li argued, can accomplish what previously required a team of specialists. A small business owner can manage marketing, accounting, customer service, and logistics with the help of specialized agents. A researcher can conduct literature reviews, design experiments, and analyze data at a scale that would have required a lab full of graduate students. A content creator can produce, edit, distribute, and monetize work across multiple platforms simultaneously.
The concept is seductive and, like all seductive technological visions, it comes with significant caveats. The "super individual" framing implies a level playing field where anyone with access to the right tools can compete with established players. But the reality of AI agents is likely to be more uneven. The individuals who benefit most will be those who already have expertise in a domain and can effectively direct and validate the work of their agents. A skilled marketing professional with an AI agent will outperform an amateur with the same agent, because the professional knows what good work looks like and can correct the agent when it goes wrong.
This is not a criticism of the technology itself. It is a reminder that tools amplify existing capabilities more than they create new ones from nothing. The spreadsheet did not make everyone an accountant. It made accountants vastly more productive. AI agents will likely follow the same pattern, and the companies that understand this — that build their agents for expert users rather than promising to replace expertise entirely — will be the ones that succeed.
The China Context
Baidu's agent strategy cannot be separated from the regulatory and competitive environment in which it operates. China has taken a markedly different approach to AI governance than the United States or Europe, emphasizing state oversight, data localization, and alignment with national development goals. This creates both constraints and opportunities that shape how Chinese AI companies build and deploy their products.
The constraints are obvious. Baidu's agents must operate within a framework of content moderation, data privacy rules, and government approval processes that are far more prescriptive than anything in the West. The company cannot simply release an agent that browses the open web, makes purchases, and communicates with third parties without extensive review and ongoing monitoring. The regulatory overhead is significant, and it slows down the pace of iteration.
But the opportunities are equally significant. China's integrated digital ecosystem — where payment, identity, transportation, commerce, and government services are deeply interconnected through platforms like WeChat and Alipay — creates an environment where agents can be genuinely useful in ways that are harder to achieve in the fragmented Western market. A Baidu agent that can book a train ticket, pay for it, and check you in using your government ID is operating in a system designed for that kind of integration. The equivalent agent in the United States would struggle with the patchwork of proprietary APIs, authentication systems, and regulatory requirements that characterize American digital infrastructure.
This is why Baidu's agent announcements, while easy to dismiss as a Chinese company's attempt to stay relevant, deserve serious attention. They represent a different model of how AI can integrate into daily life — one that is more centralized, more controlled, and potentially more effective in the near term than the decentralized, privacy-first approach favored by Western companies.
What This Means for the Global AI Race
The Baidu Create 2026 keynote should be read alongside recent announcements from Google, OpenAI, and Anthropic as evidence of a broader industry inflection point. All of the major AI companies are shifting resources from pure model research to agent development. Google's agentic AI features for Android, OpenAI's operator products, Anthropic's computer use capabilities, and now Baidu's personal assistant agents are all converging on the same vision: AI that does things, not just says things.
The competitive dynamics of this new phase will be different from the model wars. In the model era, the winners were determined by compute access, talent density, and research breakthroughs. In the agent era, the winners will be determined by ecosystem integration, user trust, regulatory navigation, and operational excellence. These are skills that favor established technology platforms with large user bases and deep integrations — companies like Google, Apple, Microsoft, Tencent, Alibaba, and yes, Baidu.
This is potentially bad news for AI research labs that do not have their own distribution platforms. Companies like OpenAI and Anthropic have built extraordinary models, but they are dependent on partnerships and API relationships to reach end users. In an agent-centric world, the companies that own the operating system layer — the phone, the browser, the cloud — have structural advantages that are difficult to overcome through model quality alone.
The Skeptic's View
Not everyone is convinced that the agent era is upon us. Critics point out that the gap between demonstration and reliable deployment is enormous. The Baidu agent that works beautifully in a keynote may struggle with edge cases, ambiguous instructions, and unexpected system changes in the real world. The history of AI is littered with products that looked transformative in controlled environments and proved frustratingly brittle in practice.
There are also legitimate concerns about autonomy and accountability. An agent that can book travel, send messages, and make purchases on your behalf is also an agent that can make expensive mistakes, send embarrassing communications, and be exploited by malicious actors. The security implications of giving AI systems broad access to personal and financial systems are profound, and the industry has not yet developed robust frameworks for managing these risks.
Baidu's response to these concerns, at least publicly, is that agent capabilities will be rolled out gradually, with explicit user confirmation required for high-stakes actions and extensive logging to enable post-hoc review. Whether this approach proves sufficient will depend on the specific failure modes that emerge as these systems scale, and the company's track record on safety and reliability will be a critical factor in user adoption.
What to Watch Next
The next six months will be telling. Baidu has committed to a broad public release of its personal assistant agent by the end of the year, and the company's performance against that timeline will signal how confident it really is in the technology. Watch for user adoption metrics, but more importantly, watch for the depth of integration. An agent that requires constant supervision is a chatbot with extra steps. An agent that genuinely reduces cognitive load and administrative burden is a new category of product.
Also watch how Western companies respond. If Baidu's agent strategy gains traction in China, the competitive pressure on Google, Apple, and Microsoft to accelerate their own agent roadmaps will increase. The agent era may not be a winner-take-all competition, but it is a competition where second place is a dangerous place to be. The companies that establish user trust and habit formation early will have advantages that compound over time.
Finally, watch the regulatory environment. Agents that act on behalf of users raise novel questions about liability, consent, and accountability that existing legal frameworks are not well-equipped to address. China is moving quickly to establish rules, but the regulatory picture in the United States and Europe remains murky. The companies that can navigate this uncertainty most effectively will have significant advantages in the global market.
Baidu's declaration at Create 2026 was bold, but it was not baseless. The model wars are indeed winding down, not because models are unimportant, but because the best models are now widely available and broadly comparable. The next battleground is execution — turning intelligence into action, thought into outcome, potential into productivity. Whether Baidu wins that battle remains to be seen. But the company has correctly identified where the battlefield lies. And in technology, that is often half the fight.