Building Search for AI Agents with Exa CEO Will Bryk

Exa was founded in 2021 by Will Bryk, who concluded while writing a history book that sufficiently good search could automate complex research and writing. The founding thesis was that Transformer-based neural systems had become capable enough to outperform Google for deep, complex queries, consistent with what Bryk calls the bitter lesson, the idea that neural systems improve continuously as more data is added. Exa released its first search engine in November 2022, ChatGPT launched roughly two weeks later, and a retweet from Andrej Karpathy generated early attention. By early 2023, developers requesting API access for AI workflows prompted Exa to pivot toward serving AI agents as its primary customer.

Bryk argues Google is optimized for billions of consumers seeking quick answers and is poorly suited for queries like finding every competitor to a company or recruiting engineers with specific backgrounds. Google's two decades of human click data are central to its ranking advantage for consumer use cases, but that data is largely irrelevant for AI agents, which removes one of Google's most significant competitive moats for the agent market. AI agents search fundamentally differently from humans: where a human wants ten results, an agent may need one thousand or ten thousand. Agents also require a controllable, non-black-box API that handles both semantic and keyword queries, supports widely varying latency requirements, and operates at reliability closer to 99.9999 percent rather than the 99.9 percent sufficient for consumer search.

LLMs have lowered the barrier to building competitive search by reducing the engineering headcount needed for re-ranking from hundreds of people to a small team. Exa has always operated with fewer than 100 employees. Even with these efficiency gains, Exa trains its own specialized re-ranking models rather than calling a general LLM, because speed and quality at scale require purpose-built systems. Bryk contends that LLMs are commoditizing faster than search because open-source models are sufficient for a large share of tasks, and that a large portion of knowledge work is fundamentally a search problem rather than an intelligence problem.

On architecture, Bryk argues model weights are wasted when used to store factual world knowledge, and that weights should focus entirely on intelligent processing while retrieval handles knowledge. He predicts future models could be under one billion parameters, highly capable at reasoning but entirely dependent on retrieval tools. Pairing retrieval with smaller models can reduce costs by up to 20 times compared to other providers by extracting only the relevant information from documents rather than passing full context to large models. Bryk ran reinforcement learning experiments comparing a Google-wrapped SERP against Exa as a search tool for agents, and RL on Exa outperformed RL on Google in both efficiency and performance, with Exa requiring fewer calls, which he attributes partly to Exa being designed to accept more complex queries than traditional search engines support.

Exa powers Devin, the coding agent from Cognition, and testing showed the integration made Devin more accurate with fewer mistakes. Bryk says search quality for coding agents has been in the equivalent of the early 2000s, and that coding agents require fresh web information to avoid errors in every line of code. He expects coding agents and general agents to eventually converge into a single everything agent. Exa serves over 5000 businesses, and Bryk says research directions follow customer needs, with customers pushing for higher quality, speed, and better information extraction.

Bryk projects the agent economy will generate roughly one trillion dollars per year in value and predicts agentic search will be a bigger business than Google search by the 2030s, with agentic search TAM exceeding Google ads revenue by 2030. He derives the estimate from the number of LLM calls, the percentage requiring search, and cost per search, noting that agents will generate millions of searches compared to humans averaging a couple per day. He acknowledges trend-based predictions may be off by a quarter but says the trajectory is clear. He identifies infrastructure as the initial bottleneck for scaling agentic search, not algorithms or data, and notes that a search algorithm optimized for the current web of roughly one trillion pages could become 1000 times more expensive at quadrillion-page scale, requiring retrieval algorithms to be fundamentally rethought.

Bryk still personally interviews every candidate at Exa and says the primary trait he looks for is passion, describing it as a fire in the eye. He argues passion and agenticism matter more now than before because agentic tools allow individuals to accomplish almost anything, making drive the key differentiator. On internal culture, he says people work on the projects most exciting to them, and he gives a company-wide speech every Monday focused on distilling projects into a memetic core with a compelling name, arguing project names matter because they ground the mission in conversations the CEO is not part of.