Silicon Valley keeps describing AI agents as the next major computing interface. Outside the technology industry, the reaction is noticeably cooler. Josh Miller, chief executive of The Browser Company, crystallized that gap when he observed that ordinary people rarely talk about an agent they use, despite claims that the technology is ready to transform daily life.
The mismatch is not evidence that agents have no users. OpenAI said in July that its Codex and ChatGPT Work agents collectively reached about 10 million weekly users. Yet that figure sits beside a broader cultural reality: “agent” remains an industry category more than a consumer desire. People generally want a task completed, not a lesson in the architecture performing it.
Capability is not the same as demand
AI companies often measure progress through harder benchmarks, longer autonomous runs and broader tool access. Those improvements matter, but consumers judge products differently. They care whether a tool is understandable, dependable and worth changing established habits for. An agent that occasionally completes an impressive sequence but behaves unpredictably during routine work may feel less useful than a narrow feature that works every time.
Miller argues that the industry has become absorbed by the technical frontier and needs to ask what people actually want. His criticism is especially notable because his career has centered on consumer software. The Browser Company created Arc, a browser with an enthusiastic following, before Atlassian acquired the business for $610 million. His point is not that autonomous software should disappear. It is that “agent” may be the wrong starting point for product design.
The 10 million-user claim also requires context. It combines two products and comes from their vendor. It demonstrates meaningful adoption among certain users, but does not establish mainstream use, retention or satisfaction.
The language may be hiding the product
The word agent covers very different experiences: coding systems, research assistants, browser operators and workflows that act across workplace tools. Combining them into one category is convenient for the industry but confusing for everyone else. A person may value automatic meeting preparation without thinking of it as an agent at all.
This naming problem reflects a deeper design issue. Products built to showcase autonomy may expose too much complexity. Users must understand permissions, supervise actions and recover from mistakes. Every additional capability expands both the potential value and the number of ways the system can surprise its operator.
Price and speed matter as well. A delegated task that consumes substantial time or computing resources can be less attractive than a familiar application, especially when the user must inspect every result afterward.
The most successful agentic features may therefore become nearly invisible. They will be presented as a reliable way to book, organize, compare or prepare something, with autonomy serving the experience instead of becoming the headline. That shifts evaluation away from what the model can theoretically do and toward completion rates, correction costs and trust.
Mainstream adoption will be earned task by task
Consumer reluctance is not necessarily permanent. Early products frequently arrive before conventions that make them approachable. Better permission controls, clearer previews and narrower areas of responsibility could turn agent technology into something people use without adopting the industry’s vocabulary.
But the burden rests with developers. Users should not have to invent a reason to care about a capability because technology leaders find it exciting. Companies need to identify recurring frustrations, design an experience around them and prove that delegation is safer and easier than doing the work manually.
Miller’s critique turns the agent debate in a useful direction. The important question is not whether models are technically capable of acting. It is whether a product gives someone enough value, confidence and control to let them act. Until that answer is obvious, the agent era may remain much larger in presentations than in everyday life.
