In May 2026 Baidu proposed a new headline number: Daily Active Agents, or DAA, offered as the agent era’s replacement for Daily Active Users [1]. The argument behind it is stronger than the metric that carries it.
Robin Li’s framing is that tokens “represent cost, not value”, because they measure input rather than output, and that a platform’s health is better judged by “the number of agents actively working and delivering results” [1]. He predicted global DAA could pass ten billion.
Behind the slogan sit products: a general agent, a coding agent whose own codebase is 90 per cent generated by itself, a digital human platform, a self-evolving planning agent, and an AI cloud sold to run them [1]. A metric launched with a product line attached is worth reading twice.
What the metric gets right
DAU measures attention. A user who chats for an hour and a user who closes the laptop while an agent finishes a report score the same, and the second one got more done. Any metric where spending more tokens looks like success is a cost metric wearing a growth metric’s clothes.
DAA at least points at delivery. “Did the agent finish the task” is a better question than “did the model answer”, and better than “how much did it consume getting there” [2].
Defining the yardstick is also a platform move, and a legitimate one. Microsoft has counted Windows in monthly active devices for years, and a billion of them was a headline, not a footnote [3]. Baidu has been early before: an NLP department in 2010, a deep learning institute in 2013, Apollo in 2017, ERNIE in 2019 [2]. Being the company that names the era is part of the business model.
Where it breaks
The definition does two jobs and only one of them is measurable. “Agents actively working and delivering results” is a count of agents plus a claim about results, and the count only covers the first half.
| Problem | Why it matters |
|---|---|
| Undefined unit | an agent can be a session, a process, a sub-agent, or a retry. One user fanning out into twenty sub-agents registers as twenty active agents |
| Counted by the vendor | the platform defines what an agent is, so it defines the number. There is no independent denominator |
| No outcome term | a count of active agents says nothing about success rate, human rescue, or cost per task |
| Not comparable | people are a shared unit across products; agents are not, so the number cannot be benchmarked |
| Architecture bias | systems that fan out into many small agents beat systems that do the work in one, on a metric that never mentions quality |
Ten billion daily agents is a prediction about a quantity whose definition belongs to whoever reports it. That does not make it false. It makes it unfalsifiable, which is worse for a metric.
Four numbers with denominators
If the goal is to know whether agents are working, these are the numbers I would put on the wall.
- Completed tasks per active user per week. The human is the denominator, so architecture cannot inflate it.
- Success rate by task class, with a human rescue counted as a failure. This is the uncomfortable one, and the only one that says whether the agent can be trusted.
- Cost and wall-clock per completed task. Two agents that both succeed are not equal if one costs ten times as much.
- Repeat use of the same workflow in week two. A one-off is a demo. The second week is the product.
The last two are not optional, because the alternative is a dashboard that counts attempts.
Why the horizon decides the ceiling
METR measures the length of software task an agent completes at 50 per cent reliability, and their published doubling time is about 201 days across the whole record, 176 days if only data from 2023 onward counts [4]. The curve is real and it is steep. It also says the reliable horizon today is measured in tens of minutes, maybe a couple of hours.
That has a direct consequence for DAA. If agents reliably finish tasks of an hour or less, then a daily count of active agents will be dominated by short tasks: lookups, edits, small code changes, one-off queries. Useful, but not the digital labour the number implies. A metric built on counts will keep reporting growth after the useful work has plateaued, because the cheap tasks are unlimited.
The two companies are shaped differently
Baidu’s history is technology first and commerce second: early on search, early on AI, late to mobile, and still looking for the application layer that pays. Its answer is to define the era and then define the metric. A cloud business sells a unit of work, so a metric counting agents serves the seller.
Xiaomi is the other shape. The phone, the car and the home already sit inside a daily routine, so “daily” is structural there and nobody needs to invent it. The agent’s job is to live inside a device people already touch, which is a harder engineering problem and an easier distribution problem. The two companies are not competing on the same axis, and their metrics should not be compared.
The split worth keeping
Design for daily, measure delivery.
The design principle is right. An agent that needs its own destination app is a novelty; the one embedded in a loop you already have, a terminal, a launcher, a phone, is daily by construction. That is a product decision, and it is the interesting half of what Baidu is saying.
The metric is not ready. Until DAA says what an agent is and divides by something the vendor does not control, it is a slogan with a number attached: useful inside a company to point teams at completion, unusable between companies as a claim about productivity.
References
[1] Baidu Advances Agent Portfolio to Embrace the Agent Era, Champions Daily Active Agents as Key Metric: the DAA proposal, the tokens-are-cost quote, the ten-billion prediction, and the product line launched with it.
[2] Baidu wants daily active agents to become AI’s next value metric (KrASIA, from 36Kr): the DAU-versus-delivery argument and Baidu’s earlier bets on NLP, deep learning, Apollo and ERNIE.
[3] Windows 10: Powering the world with 1 billion monthly active devices: the precedent for a vendor choosing the unit and turning it into the headline number.
[4] Measuring AI Ability to Complete Long Software Tasks (METR): the 50 per cent time-horizon method and the doubling times of roughly 201 days across the record and 176 days since 2023.