Configurable Retrieval for AI Agents
Will Bryk, CEO of Exa, on building search for AI agents
This reveals that Exa is building search as configurable infrastructure, not as a one size fits all consumer box. Google optimizes for a universal experience where every query feels equally fast and simple. Exa is instead treating search more like cloud compute, where a customer can decide when a hard query deserves more work, more cost, and better recall, especially when an AI agent is searching for raw inputs rather than a human looking for one obvious link.
-
In practice, Exa is trying to route between semantic and keyword search depending on the job. A query like William Bryk LinkedIn is cheap and easy, while a query like startups building futuristic hardware needs meaning based retrieval and deeper filtering. That is the core philosophical break from Google style search.
-
That flexibility matters because Exa customers often use search in automated pipelines, not as a consumer destination. One user runs 5,000 searches a day, asks for up to 10,000 results per query, pulls full page text, and feeds the output into downstream agents. For that workflow, control over result depth matters more than a fixed 400 millisecond response.
-
The market is splitting into two product shapes. Exa is strongest when customers want many raw results and full text to power their own agents. Parallel is stronger for longer research tasks that return a finished synthesis. That makes Exa look less like a better search page and more like the retrieval layer inside agent software.
Search for agents is moving toward tiered retrieval, where simple lookups stay cheap and complex research queries trigger more compute, more filtering, and more structured output. As agent traffic grows faster than human search traffic, the winners will be the providers that let developers tune that tradeoff directly and build whole workflows on top of it.