Search In the Age of Agents

Information steers civilization.

That’s how Exa (opens in new tab) CEO and co-founder Will Bryk describes the stakes of the company’s work. It’s a mission that’s motivated him since 2021, when he and co-founder Jeff Wang set out to build a better Google. The market for agent search did not yet exist, but the founders’ conviction did.

Today, that founding ambition has expanded into a broader question: can better information architecture improve the way we live and work?

Search needs to be reinvented for agents. The knowledge captured inside a model is out of date as soon as the model reaches production. To do useful work, an agent needs access to current information and all the relevant context for the task. As agents take on more work, the volume and depth of their searches will grow with them.

That makes the quality of the underlying information more consequential than ever. If technology advances faster than our ability to find, evaluate, and coordinate around reliable information, the consequences extend far beyond the quality of a search result.

We believe high-quality information is the key ingredient in reliable, high-performing AI systems. That is why we at Felicis are proud to invest in Exa as part of its Series C and to support Will, Jeff, and the entire Exa team.

Exa was born out of Will’s natural curiosity. He wanted to write a history book, but soon found that existing search tools were poor at surfacing the best, most relevant, and most complete information. The time he assumed he’d spend writing was quickly consumed by filtering through poor search results.

The search problem lingered with Will as he worked on real-time AI products at Cresta (opens in new tab), so he reached out to his former Harvard roommate, Jeff Wang. While Jeff had moved onto the data and web infrastructure team at Plaid (opens in new tab), he and Will had already played around with building a search engine together in the past.

Agents Do Not Want Links

Agents search differently from people. A person might enter a three-word query and navigate a page of links. An agent starts with a much richer instruction and needs information it can use directly. Consider an agent evaluating a stem-cell company. It may need to find recent scientific research, clinical developments, patents, team history, and comparable companies, then weigh those sources together. It does not want a page of blue links. It needs the underlying information.

Exa is built for the new agentic era (opens in new tab). The team has built its own search index and the infrastructure around it, giving Exa control over how the web is crawled, retrieved, and returned to agents. Exa can interpret detailed natural-language queries, retrieve the contents of pages rather than simply rank URLs, and return relevant material in a form an agent can use.

The value to the end user isn’t more results. It’s a better answer. If a search misses the decisive source, an agent’s output can sound confident and still be wrong. More complete retrieval reduces those blind spots. More precise and current information improves accuracy. Speed matters because latency compounds when an agent searches dozens of times to complete a single task.

Agent search is easy to demonstrate and difficult to evaluate. Production traffic includes ambiguous questions, obscure subjects, changing information, and queries that no benchmark designer anticipated. No surprise, then, that public benchmarks fall short.

Exa goes beyond them by using representative query distributions and private held-out sets.

A Modern Bell Labs

When we first met Will in 2024, the conversation quickly moved beyond the product. We discussed search architecture, the history of information, and what agents would eventually require from the web. It immediately became clear that Exa’s ambition extends beyond building a better retrieval system. As agents increasingly consume information on behalf of users, publishers and creators need a reason to continue producing valuable work. Exa Connect (opens in new tab) creates a more direct relationship between AI products and the publishers whose information they use. It aligns incentives between quality information distribution and retrieval, rather than maximizing the time someone spends on a search results page looking at ads.

Will and Jeff want to build a modern Bell Labs, a place where unusually strong researchers and builders can develop foundational technology that other laboratories, companies, and products rely on. The commercial traction bears this out: Exa already powers search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com, and over 500,000 developers.

We were drawn to the discipline behind that ambition. The team measures technical progress against the outcomes its customers care about: whether their agents find the right sources, return more accurate answers, and do so within the speed and cost requirements of their products. That combination matters when the goal is as large as building the information infrastructure for a new class of software.

The best search engine for agents may be one most people never realize they’re using. All they’ll see is better understanding, clearer decision-making, and more of the work getting done.

We recently sat down with Will to chat about all this and more. Watch the full conversation below.

Transcript

Will Bryk, Co-founder and CEO, Exa
Tobi Coker, Partner, Felicis
Peter Deng, General Partner, Felicis

[00:00]Will Bryk

If you have a civilization whose technology is advancing extremely rapidly, but its ability to coordinate and its information tools do not improve in quality, that is a very bad combination. That was very motivating to me. Information steers civilization.

[00:28]Tobi Coker

Will, thanks for coming in. We are excited to have you here to talk about all things Exa. This relationship has been years in the making. I was looking back at our first email, which went out on March 12, 2024.

[00:45]Will Bryk

Very cool.

[00:46]Tobi Coker

We always say that the best founder relationships develop over time, so we are excited that we finally got to partner together. I remember you talking about the vision for Exa, which at the time was enabling search for enterprises. This was before agents, and you had already laid out that vision. It has obviously been a meteoric rise since then.

[01:09]Will Bryk

We have been doing Exa for five years. In the early days, in 2021, we were not thinking about building search for agents. That came two years in, in 2023, when people started building the first AI products and needed a search API. Back in 2021, we were thinking about ourselves. Humans need high-quality information, just like AI systems do, but we did not realize the connection at the time. Before Exa, I was trying to write a history book and go deep into the topics I covered.

[01:41]Tobi Coker

You were trying to write a history book?

[01:43]Will Bryk

Yes. I never thought of myself as a history person. I was a math and science person, and then I started reading history books and thought, this is insanely cool. I figured that by writing a book, I would also learn a ton of history. Doing research with Google, I found it was really hard to get to the bottom of anything. I was excited about the book, but then I realized that search could be solved. This was even more important because we could build something that impacts billions of people, not just millions of readers.

[02:11]Peter Deng

What you are building is truly amazing and generational, but few people understand it as deeply as we do. When I was starting my career, people who were considering joining Google would hear, ‘Why are you joining Google? Search is a solved problem. Everyone just goes to Yahoo.’ That seems laughable now, but people are saying something similar today. They dismiss search as solved. What are they getting wrong?

[02:42]Will Bryk

People do not realize that there is a new user of search: AI systems. They need a completely different system. Even calling it search makes you think of the old days of Google. I have started calling it information. Our goal is perfect information, not just search.

[02:56]Will Bryk

What does perfect information mean? These AI systems can consume information in a way humans cannot. If you are building the optimal product or agent, you want as much high-quality information as possible. If you built that same system for a person, they would not know what to do with all of it.

[03:21]Peter Deng

There is a big difference between listicles and blue links and what Exa is trying to do. AI needs a great deal of high-quality information because it can actually read it all. Break that down for us. How does an agent search, and how is it different from the way a person searches?

[03:39]Will Bryk

The biggest difference is that humans are lazy and agents are not. A person types a few keywords, gets a list of links, clicks one or two, consumes some information, gets bored, and moves on. Google was optimized for that kind of user.

[03:58]Will Bryk

An AI can instantly produce a paragraph explaining exactly what it wants, including any filters it needs. It can issue complex queries and consume dozens, hundreds, or even thousands of results very quickly. Search for AI should handle those complex queries and filters and return as much useful information as the agent needs.

[04:28]Will Bryk

There is another important difference. When you build search for people, you want every link to match the query because the person might click result one or result four. An AI consumes the results together, as one body of information. You want diversity within that set of results, not the same answer repeated over and over.

[04:45]Peter Deng

These systems can also be manipulated by misinformation. They tend to believe what they see. Because agents are now taking actions, you want the information behind their reports to be as accurate as possible.

[04:59]Peter Deng

Last night was my fantasy football draft. I tried a new strategy using Claude Opus with Exa connected. I described my strategy and told it to go. It ran all of these Exa searches. I will tell you at the end of the season how it did, but it was a completely different way to consume information. I did not have to read every article. The agent could ingest all of it and apply my strategy.

[05:29]Will Bryk

That is a good example because you care about fantasy football and you want to win. Why would you not give your agent access to the best information? It does not even cost very much.

[05:41]Peter Deng

Exactly. Because I could describe the strategy precisely, it could research injury reports, quarterback and wide receiver relationships, and everything else that mattered. That is now possible because of Exa and Claude together.

[05:55]Tobi Coker

When we were first considering the investment, I was looking at a stem-cell company. I tried to use Google Scholar to find the relevant research, but I could not get what I needed. This was around the time you were transitioning from Metaphor to Exa. I used the research search product you had then, and in about 500 milliseconds it returned more relevant research links on the company than I had found through multiple Google searches.

[06:28]Tobi Coker

Help us walk through the numbers. We believe the number of agents will grow exponentially from here. Help us do the math, project out what that future could look like, and explain how you build Exa to scale with that infrastructure.

[06:49]Will Bryk

Some numbers help illustrate what is coming. Google has two to three billion users who search a few times a day, which is roughly 15 billion queries daily. In a few years, it is reasonable to imagine every person having at least one agent. A personal assistant agent will not run only a few searches a day. It could run hundreds or thousands, because one interaction can trigger many searches and a person may have hundreds of interactions.

[07:36]Will Bryk

That could mean hundreds of millions of queries per second, around a thousand times more QPS than Google handles today. Serving current-quality search at that scale is already an extraordinary engineering problem. Every part of the system has to operate at that volume.

[08:00]Will Bryk

At the same time, agents demand higher quality and precision than people do. Quality comes from two places: the index and retrieval. You want to gather as much valuable data as possible, not just the public web, and keep it extremely fresh. As the index grows, retrieval becomes more difficult. You have to search across trillions, and eventually quadrillions, of items as quickly and cheaply as possible. That is a very hard and underexplored machine-learning problem.

[08:57]Tobi Coker

You built the early web crawler and models yourself. Seeing the system at its current scale must give you an appreciation for how difficult the problem is.

[09:08]Will Bryk

It is crazy to see that some of the names I originally chose are still there, while the systems are now hundreds of times larger. The GPU cluster is about 100 times larger, and the amount we crawl is probably about 100 times larger. I hope my original code is not still there because it was pretty bad human code. All code should basically be written by well-informed AI.

[09:34]Tobi Coker

You have also released benchmarks and evals. How much weight do you place on them? What are you optimizing for, and why are those evals important?

[09:49]Will Bryk

Evals are extremely important, arguably the most important thing. In agent workflows, once you know exactly which hill you are climbing, it becomes much easier to climb it. We have hundreds of internal evals as well as some public ones.

[10:01]Will Bryk

Public evals have a weakness: once the questions are public, systems can be tuned to them. A strong evaluation program needs a private, held-out set. You also have to ask whether the distribution of queries in an eval represents anything that matters. Some BrowseComp queries are deliberately unusual and very different from the questions our customers ask. For our internal evals, we look at what customers actually care about and build evaluations that capture it.

[10:33]Will Bryk

I have not done a good job of explaining what perfect information does for the world. Technology is advancing extremely fast. We are moving extremely quickly, but if we do not steer ourselves with information, we will end up in a very bad state. It is not sustainable. That has been motivating to me for five years because I think we can help solve it.

[10:52]Will Bryk

You also need an organization with the right incentives. Most of the world’s major information tools are not perfectly aligned with a high-quality information society. Mainstream news has an incentive to be provocative. Social media and traditional search monetize human attention through advertising. These products have done a great deal of good, but their incentives are not perfect.

[11:47]Will Bryk

Exa and this new generation of companies are different. I wake up to customers saying they would pay us more if search improved in a particular way or if the information were higher quality. We are directly incentivized to improve information quality.

[12:05]Will Bryk

If we move from a world in which people consume most information and advertising is the dominant business model to an agent economy in which the business model rewards high-quality information, incentives change. Data and news providers have a reason to create information that is accurate, distinctive, and valuable, rather than simply provocative.

[12:47]Tobi Coker

You recently launched Exa Connect, which lets publishers set their own price. That feels like a step in this direction. Why did you build it, and how do you see it developing as agents become more common?

[12:59]Will Bryk

We want to enable that world and be one of the big marketplaces. Exa Connect is a simple idea. We already have many agents using Exa. A provider that wants to reach agents should not have to create 5,000 separate relationships. It can partner with Exa, and we can route the data it chooses to expose to those agents.

[13:25]Will Bryk

For Exa, that means we can combine information from many providers with the public web. Agents get access to more useful data, and providers gain a new distribution channel and business opportunity.

[13:42]Peter Deng

You wrote a manifesto before the company took off. What was in it?

[13:49]Will Bryk

I wrote a lot in the early days. One document laid out questions and answers about how the company should be structured. In 2021, saying that we were going to build a better Google sounded insane. I remember telling my dad, and he said, “You are insane, but I fully support you.” At the time, Google and search felt synonymous. You had to dream about what perfect information could mean. Even today, that is difficult for people to imagine. I also wrote a post called “High-Quality Information Society.”

[14:31]Peter Deng

In that document, you talked about building a modern Bell Labs. What does that mean?

[14:37]Will Bryk

The original idea was that we would be a research organization exploring information in new ways. Bell Labs tried many things without knowing which ones would work, and that curiosity produced tools like the transistor. We are not there yet, but it remains a guiding principle. We want to become large enough to have the space and capacity to explore many interesting problems in information.

[15:07]Will Bryk

We are exiting the first phase of Exa, which was essentially to build Google internally and make it available to the world. We have basically achieved that. What I did not realize in 2021 was that we would also become an enabler of other Bell Labs. We started by thinking we might build a platform, but we became infrastructure. Many companies and AI labs now use Exa underneath their work on new science, mathematics, and other ambitious problems. Enabling those Bell Labs is pretty special.

[15:45]Tobi Coker

What do you look for in the people you hire?

[15:51]Will Bryk

I think a lot about culture and about making sure great people feel good at the company. In one sense, we stumbled into the culture we have. We hired people we wanted to spend time with, who were smart and passionate. Those early employees set the culture for everyone who followed because eventually they began conducting the technical and go-to-market interviews themselves.

[16:25]Will Bryk

The most important part of culture is the people you hire. The excitement and the mission are mixed in, too. It may be hard to have an unexciting culture when you are trying to do something this ambitious and genuinely believe in it. I would describe Exa as really smart people who really care.

[16:48]Tobi Coker

Tell us about the shirt you are wearing. It says 10 to the 18th and 10 to the 100th.

[16:53]Will Bryk

Ten to the 100th is a googol, and 10 to the 18th is an exa. The shirt says Exa is greater than Google, which is mathematically wrong. People stop and look at it because it is provocative. The best marketing makes someone say, “Wait, what? You are wrong,” and then talk about it.

[17:14]Tobi Coker

It got me to ask the question. You also had a billboard.

[17:16]Will Bryk

We put the same idea on a billboard. We knew the billboard itself was not the point. Someone would share it, and that post would travel. One post about it got 20 million views. We think a lot about creating something people want to share or discuss, while balancing that against being provocative just for the sake of it.

[17:46]Tobi Coker

Will, thank you for coming. This was an electric conversation. Exa to the stars. We really enjoyed it.