Google used I/O 2026 to make a bigger claim about Search: it does not want to be only the place people type questions and click links. It wants to become the place where users start and supervise AI agents.
In its Search announcement, Google said AI Mode has passed 1 billion monthly users one year after launch, with queries more than doubling every quarter. The company is now upgrading AI Mode with Gemini 3.5 Flash and rolling out what it calls the biggest Search box upgrade in more than 25 years.
The shift: from results page to task surface
The most important change is not another chatbot answer. It is the idea of Search agents that can be created, customized, and managed inside Search.
Google’s first examples include background information agents, agentic booking, shopping flows, and generated mini apps. The company also says users will be able to move from an AI Overview into AI Mode follow-ups, with context carrying over between the summary and the conversation.
That turns Search into a task surface. A query can become a tracker, a dashboard, a shopping assistant, a booking workflow, or a small generated interface instead of a list of blue links plus an AI paragraph.
WIRED’s I/O recap framed the broader keynote the same way: Google is pushing agents into Search, Gmail, YouTube, Docs, Chrome, and other major services. CNBC reported that Google also announced Gemini Spark, a general-purpose agent meant to act across connected apps under user direction.
Why this matters for publishers and software companies
Search becoming more agentic changes the traffic bargain. If Google can answer, compare, book, track, and generate an interface inside Search, fewer user journeys need to leave Google’s own surface.
That is useful for users when the task is simple. It may be painful for publishers, marketplaces, and software tools that depend on Search as a referral engine. A booking agent or shopping flow can compress many pages of research into one controlled experience. The question is who gets visibility, who gets attribution, and who gets the transaction.
Google is trying to answer a different business question at the same time. Investors want proof that huge AI spending creates product leverage. Search is the obvious place to show it because it already has distribution, intent, ads, commerce, and user habit. If AI Mode makes people search more, as Google says, the company can argue that AI is defending the core business rather than replacing it.
Developers are part of the same agent push
The I/O developer keynote points in the same direction. Google’s developer recap described a move from AI that assists users to agents that navigate complex tasks across workflows.
Google announced upgrades to Antigravity, an Antigravity CLI, Managed Agents in the Gemini API, and tighter AI Studio integrations. The message is clear: agents are not just a consumer Search feature. Google wants developers building on the same model of task automation, remote sandboxes, and orchestrated workflows.
That matters because the winning agent platforms will not be judged only by model quality. They will be judged by tool access, safety controls, deployment paths, identity, billing, and whether developers can plug agents into real work without building the plumbing from scratch.
What to watch next
The first test is quality. A bad list of links is annoying. A bad agent that books the wrong thing, summarizes from weak sources, or hides important options is more consequential.
The second test is control. Users need to know when an agent is researching, when it is acting, what sources it used, and how to undo or stop it. Background agents are useful only if they feel accountable.
The third test is the web economy around Search. If generated interfaces and agentic workflows keep more activity inside Google, publishers and software companies will push for clearer attribution and commercial terms.
Google’s I/O message is practical, not subtle: the search box is becoming the front door for agents. The opportunity is a faster, more useful internet interface. The risk is that the open web becomes another input layer for a platform that increasingly completes the job itself.
