Google CEO Sundar Pichai’s description of search as a future “agent manager” made headlines this week after a 1-hour interview with Patrick Collison, CEO of Stripe.
As reported by SEJ’s Roger MonttiPichai described a version of search in which users have “many threads running” and perform tasks rather than browsing results.
But the interview covered much more than just that one quote. Throughout the conversation, Pichai laid out a timeline, identified obstacles slowing adoption, described how he already uses an internal agent tool, and confirmed the infrastructure constraints that limit how quickly this vision can be implemented.
Here’s what the rest of the interview reveals to research professionals.
How Pichai’s language intensified
The “agent manager” line didn’t come out of nowhere. Pichai’s language on the future of research has become more precise over the past 18 months.
In December 2024, he told an interviewer that research “would change profoundly in 2025” and that Google would be able to “tackle questions more complex than ever”.
In October 2025, during Google’s third quarter earnings call, he referred to this as ” “Moment of expansion for research” and reporting that AI queries had doubled quarter over quarter.
In February 2026, he reported Search revenue reached $63 billion in fourth quarter 2025 with growth accelerating from 10% in the first quarter to 17% in the fourth quarter, attributing the increase to AI capabilities.
Now, in April, he’s putting a label on it. Not “search will change” or “search is expanding”, but “search as an agent handler” where users perform tasks.
Each time, the language moved from abstract to concrete, from prediction to description.
The 2027 inflection point
Collison asked Pichai when a fully agentic business process, like automated financial forecasting without human intervention, might happen at Google. Pichai pointed to next year.
“I certainly expect that, in some of these areas, 2027 will be a significant inflection point for some things.”
He added that non-technical workflows would see “pretty profound” changes in 2027, noting that some groups within Google already work this way.
“There are certain groups within Google that are evolving more deeply, and so for me a big task is how to spread this to more and more groups, especially in 2026.”
He also recognized that younger, AI-native companies have an advantage in adopting these workflows, while larger organizations like Google face reskilling and change management challenges.
The intelligence overhang
One of the most useful parts of the interview didn’t come from Pichai. This is Collison’s description of what he calls “intelligence overhang,” the gap between what AI can do today and how much organizations actually use it.
Collison identified four barriers that slow adoption, even when models are successful. The first is to encourage competence. Achieving good results with AI takes practice, and most people in organizations have not yet acquired this skill.
The second is the company-specific context. Even an experienced teleprompter needs to know which internal tools, datasets, and conventions to reference. The third is data access. An agent cannot answer “what is the status of this transaction?” » if he cannot access the CRM or if permissions are blocking him. The fourth is role definition. Job descriptions, team structures, and approval workflows were designed for a world without AI colleagues.
Pichai agreed with that assessment and said Google faced the same challenges internally.
“Identity access controls are like real hard problems and so we’re working on those things, but those are the key things that also limit the spread to us.”
He described how Google’s internal agent tool, which he called Antigravity, is already changing the way he works as a CEO. He said he was interviewing her to get quick information on product launches.
“Hey, we launched this thing, what did people think about it? Tell me the five worst things people are talking about, the five best things people are talking about, and I’ll type that.”
This is a real-world example of the agent manager concept in action today within Google. Pichai uses search as a task execution tool, not a link referral tool. The gap between that internal experience and what’s available to external users is part of what Google is working to fill.
For SEO teams and agencies, information overload is worth thinking about on two levels. There is a surplus in your own organization, where AI tools could do more than they currently do. And there’s the advantage on Google’s side, where the models are already capable of agent-style search, but the product hasn’t fully delivered it yet.
What is blocking the timeline
Pichai confirmed that Google’s capital spending in 2026 will be between $175 billion and $185 billion, correcting the $150 billion figure cited by Collison. That’s about six times the $30 billion Google was spending before current AI development.
Asked about bottlenecks, Pichai identified four constraints in order.
Platelet production capacity is the most fundamental limitation. Memory provisioning is “certainly one of the most critical constraints today.” Permitting and regulatory delays for the construction of new data centers are a growing concern. And critical supply chain components beyond memory add additional pressure.
“It’s not possible for the major memory companies to significantly improve their capacity. So you have these short-term constraints, but they ease up as you go out.”
He said these constraints would also lead to efficiency gains, predicting Google would make its AI systems “30 times more efficient” while increasing its spending.
He also noted that he personally dedicates an hour each week to reviewing compute allocation at a granular level across Google teams and projects.
What this means for research professionals
Pichai’s description of search as an agent manager changes the question SEO professionals need to ask about their work.
In an outcome-based search model, the goal is to rank. In an agent-based model, the goal is to be useful to a system that accomplishes a task. These are different problems.
Consider what agent-performed research looks like in practice. You tell the searcher to find a plumber, check reviews, confirm availability for Saturday morning and make an appointment. The agent does not return ten blue links. It leverages structured business data, review platforms, and reservation systems to complete the job. The companies chosen are those whose information is accurate, structured and accessible to the agent. Those with outdated schedules, no booking integration, or thin review profiles don’t appear.
The same pattern applies to e-commerce. A shopper says, “find me running shoes under $150 that fit flat feet and can arrive by Friday.” » An agent capable of accomplishing this task needs product data, inventory availability, shipping estimates, and compatibility information. Sites that provide this data in structured, machine-readable formats are now part of the agent’s toolbox. Sites that bury it in JavaScript-rendered pages or behind login walls are ignored.
If an agent can synthesize a response from five sources without sending the user to any of them, what is the value of being one of those five sources? This depends entirely on whether the agent cites you, links to you, or treats your content as raw material without attribution.
This matches the changes we’re seeing in AI mode. Google reported during his Fourth Quarter 2025 Earnings Call that AI mode queries are three times longer than traditional searches and frequently ask follow-up questions.
The 2027 calendar also matters. If non-technical company workflows begin to become agent-based next year, the companies providing the information and services these agents derive will need to be structured for machine consumption, not just human navigation. Structured data, clear APIs, and accurate business insights become infrastructure, not assets.
The measurement gap
Pichai’s insistence that AI research is not zero-sum deserves more scrutiny than it usually gets.
He constantly made this argument. In October 2025, he called it “moment of expansion”. In February 2026, he said that Google I hadn’t seen any evidence of cannibalization.. In this interview, he compared it to YouTube which thrives despite TikTok.
But total query growth and traffic on each site are different metrics. Google may be right that more people are searching more often, while individual publishers and businesses see less referral traffic from those searches. Both things can be true at the same time.
Google did not share outbound click data from AI mode. Until Google provides this data, Pichai’s “expansionist” claim is an assertion and not a verifiable fact. Search professionals should track their own referral traffic trends independently rather than relying on Google’s characterization of the entire market.
Looking to the future
Pichai’s language in this interview goes further than what Google has said publicly before. Previous statements described AI research as an evolution. This one puts a clearer label on Google’s search direction. Research as an agent manager is a product vision.
The timeline he presented, with 2027 as the inflection point for non-technical agent workflows, gives you a window. How Google monetizes the tasks agents complete, whether agents cite sources or simply use them, and what visibility even means in an agent manager model are all open questions that will need answers before 2027.
Google I/O 2026 is scheduled for May 19-20 and will likely provide more details on how these features will be delivered.
More resources:
Featured Image: PJ McDonnell/Shutterstock
