Building Search for AI Agents with Exa CEO Will Bryk

Exa was founded in 2021 after Will Bryk concluded that Transformers had become capable enough to build a fundamentally better search engine than Google. His motivation came partly from writing a history book covering five thousand years of daily life, a project that exposed how badly Google performs on deep or complex queries. Bryk argues Google is optimized for consumer use cases and human clicks, not for tasks like finding every competitor to a company, recruiting engineers, or understanding life in the Roman Empire in 100 AD.

The original architectural thought experiment was running GPT-3 over all trillion documents on the web for every query, which Bryk estimated would cost roughly ten billion dollars per query, framing the entire engineering challenge as an optimization problem. Exa built from scratch rather than adapting existing search infrastructure and has kept its team below one hundred people. The first search engine launched in November 2022, Andrej Karpathy retweeted it, and Bryk describes it as the first moment people felt it was possible to find things beyond Google. ChatGPT launched approximately two weeks later, and by early 2023 developers were requesting API access, at which point Bryk realized the search engine built for technically deep personal use was nearly identical to what AI agents needed.

Bryk argues agents differ from human searchers the way humans differ from sloths. Agents have effectively infinite time, make complex queries rapidly, demand perfect output, and want not ten results but potentially thousands or tens of thousands, because missing a critical piece of information is unacceptable. He maintains a list of approximately twenty ways agents and humans differ as search users. Google's interface, which he describes as uncontrollable and untoggleable, is the opposite of what agents require, and the twenty years of human click data Google accumulated provides little advantage when the end consumer is an AI rather than a person.

Bryk contends that LLMs have changed the competitive dynamics of search infrastructure. A single engineer can now build a re-ranker that previously required hundreds of people at Google, though Exa still trains its own custom re-ranking models rather than simply calling a general LLM, because speed and quality require purpose-built systems. He also argues that improving model intelligence can partially substitute for the large behavioral signal Google has accumulated, helping newer entrants close the quality gap. Exa ran reinforcement learning experiments comparing a Google SERP wrapper against Exa and found that RL on Exa produces fewer search calls and higher performance. Bryk argues that big ideas from LLM training, including pre-training, post-training, and RL on embedding models, apply equally well to training search models.

Exa powers Devin, the Cognition coding agent, and testing showed it makes Devin more accurate with fewer mistakes. Bryk describes coding agents as having been in the dark ages of search quality despite their raw intelligence, because they lack access to fresh and comprehensive information. On cost, Bryk says Exa can save customers up to twenty times compared to other providers by extracting only the most relevant information from documents rather than returning full pages. He predicts models could reach below one billion parameters while remaining highly intelligent if retrieval handles factual lookup, and says this architectural shift toward smaller models that offload knowledge to cheap external tools is already beginning.

Bryk defines search as far broader than the 1998 text-box-and-keywords paradigm. He argues political polarization is fundamentally a search problem caused by people receiving inaccurate or misleading information, and that loneliness is a search problem because people cannot efficiently find others with compatible interests. He estimates the agentic economy could eventually generate a trillion dollars a year in value and suggests a meaningful portion could be distributed back to the data sources that make search possible.

Bryk predicts agentic search will be noticeable as a major trend by the end of 2026 and will be a bigger business than Google search by the early 2030s. He derives this by estimating the number of LLM calls, the percentage requiring search, and cost per search, then extrapolating the trend. His core premise is that while humans make a couple of searches per day on average, individual agents will generate millions of searches each. The near-term bottleneck for scaling agentic search is infrastructure rather than intelligence or data, with vector database throughput and compute as current constraints. A longer-term bottleneck will be data not recorded anywhere on the web, including satellite imagery and knowledge held only in people's heads.