Baidu proposes DAA (Daily Active Agents) to redefine AI, shifting focus from model capability to app engagement, but the metric is questionable and needs scrutiny of model power and platform openness.
Robin Li introduced a new term at Baidu Create 2026: DAA, Daily Active Agents. Frankly, this might be the most ambitious—and most self-serving—concept to come out of China's tech sector this year.
It's ambitious because Baidu is trying to redefine the yardstick for measuring the AI industry with a single metric. In the past, how did we judge whether a large language model was good? We looked at parameter size, benchmark scores, and API call volumes. Baidu says those aren't enough—what matters is whether agents are actually being used. Shifting from "model capability" to "application engagement" isn't a wrong approach in itself. It's like how in the smartphone era, we don't obsess over the chip's nanometer size (though it matters); we care about which apps you open daily. At its core, DAA hands the power of evaluation from the lab to the users—does anyone click on your agent, and how many come back every day?
It's self-serving because Baidu knows exactly how favorable this metric is to its own position. Wenxin Yiyan's daily active user numbers have been stuck in an awkward range, and compared side-by-side with ChatGPT and DeepSeek, the figures don't look good. But agent counts? Baidu has spent over a decade in search and holds the largest base of small and medium-sized enterprise customers—a barbershop can register a "smart hairstyle advisor," a restaurant can launch a "smart ordering assistant," and even if each agent only has double-digit daily actives, collectively it makes for a compelling story. DAA isn't a metric that grew from nothing; it's a chessboard Baidu has redrawn using its existing advantages.
But the cleverness of this concept is that it's hard to directly call it wrong. Agents are indeed the key form of AI deployment. Anthropic has been pushing the Agent concept, and Google's Project Mariner and Astra are also heading in that direction. The entire industry is moving from "chatting" to "doing"—from conversational AI to agents that can execute tasks for you. Baidu hasn't strayed off course in this narrative. The question is whether Baidu's "agent" and Anthropic's "Agent" are the same thing.
Anthropic's Agent completes a multi-step task in a single call—"help me send the meeting schedule emails to everyone for next week, and book a conference room while you're at it"—it autonomously plans its toolchain and automatically retries when a step fails. Baidu's agent, on the other hand, looks more like a WeChat official account customer service bot wrapped in an AI shell—"Hello, I'm the smart advisor for XX store, how can I help you?" There's a fundamental difference: the former is capability spillover, the latter is traffic monetization.
These clearly aren't on the same level. But Baidu probably doesn't care. Baidu's battleground has never been at Silicon Valley's table. Its ecosystem is the existing inventory of the Chinese-language internet: tens of millions of small and medium-sized merchants, vast amounts of web content, and the channel advantages of a search engine. In this ecosystem, a "good enough" agent has more commercial value than a "cutting-edge" Agent. Baidu is fighting a war of proliferation, not one of benchmarks. The reason DAA holds up as a metric is that it aligns with Baidu's DNA—connecting people with information (now upgraded to connecting people with services)—rather than competing on model ceilings.
But there's one question that can't be sidestepped: does Baidu have the right to define industry standards?
My take is: no. In the summer of 2026, the yardstick for the AI industry still rests on the boundaries of large model capability. The premise for DAA activity is that the underlying model is strong enough—if an agent gets an answer wrong once every three calls, then high daily actives are just a false boom. Wenxin Yiyan 4.5 has caught up to the top tier after several iterations, but it hasn't pulled ahead. If the underlying model is merely a qualified follower, then the ceiling DAA can prop up is limited. Baidu can say "lots of people use my agents," but if it doesn't clarify whether those users are satisfied afterward and whether their problems were actually solved, then DAA is just a vanity metric.
What's more concerning is this: when Baidu makes DAA its core KPI, how will the entire organization respond? Internal teams will rush to inflate agent numbers—splitting one feature into ten agents, breaking apart what a single Agent could handle, just to pad the denominator. This is the classic trap for any platform company pushing a "new metric": once a metric gets tied to resource allocation, behavior gets distorted. Back in the day, Baidu pushed MIP (Mobile Instant Pages) with good intentions, and it ended up becoming a textbook case of search teams' KPIs hijacking user experience. Will DAA repeat that mistake?
My stance on DAA is: half endorsement, half wait-and-see.
What I endorse is the direction—moving from model capability to application deployment, from technical metrics to user metrics. That's the inevitable trend as the AI industry transitions from the "building nuclear bombs" phase to the "using nuclear power to generate electricity" phase. Baidu's pioneering move to propose this shift in the Chinese-language context shows strategic foresight. An agent ecosystem with hundreds of millions of daily actives indeed holds more commercial imagination than a model with the top benchmark score.
What I'm waiting to see is execution—can Baidu's organizational capability keep up with this narrative? As a company where the C-side does C-side work and the B-side does B-side work, with entrenched internal interests and factions, history has proven time and again how underwhelming its strategic execution can be. An agent ecosystem needs an open, low-barrier, developer-friendly platform, not an ad network that requires layers of approval to join. Baidu's track record with developers—from Xiongzhanghao to Baijiahao—hasn't earned much trust.
Robin Li's line at the keynote—"DAA is the DAU of the AI era"—sounds polished, but anyone can sound polished. Whether DAA becomes as widely adopted a concept as DAU depends on three things: whether Wenxin Yiyan's model capability can keep evolving, whether Baidu's agent platform can truly open up, and—most importantly—whether Baidu can restrain its instinct to monetize agents and turn the ecosystem into an advertising system.
If all three come together, DAA will be the opening chapter of Baidu's comeback. If not, it's just another flashy denominator in a PowerPoint deck.