William Bryk (@WilliamBryk) and Jeffrey Wang (@jeffzwang) launched Exa Agent on June 16, turning Exa from an AI-native search API into a higher-level research endpoint for developers who want structured web work without building the orchestration layer themselves.
That shift is consistent with the company Bryk and Wang have been building since their Harvard days. Y Combinator lists Exa as a Summer 2021 company founded in 2021, with Bryk as co-founder and CEO and Wang as founder. Bryk says on YC's profile that he studied computer science and physics at Harvard, did machine learning research, led the robotics club, worked at Cresta, and left to start Exa after concluding that the future of civilization depends on the quality of information people consume. Wang's YC bio says he studied computer science and philosophy at Harvard and worked at Plaid from 2019 to 2022.
Exa Agent is the product version of that thesis for enterprise software builders. Instead of asking developers to stitch together search, crawling, model calls, retries, citations and schema validation, Exa is packaging the workflow behind a single endpoint. The company says the agent is designed for deep research, list-building and entity enrichment, and that larger jobs are split into subtasks handled by subagents that research different domains in parallel.
The product is priced like infrastructure, not a consumer research app
Exa's launch post frames the release as frontier web research at a fraction of the cost. That claim is Exa's, not independently verified, but the pricing is concrete. The Agent docs list fixed effort modes at $0.012 per request for minimal
, $0.025 for low
, $0.10 for medium
, $0.50 for high
and $1.00 for xhigh
. The default auto
mode scales compute and tool use to the task. Exa also charges $0.005 per search tool call, with separate contact-enrichment fees.
The design decision matters because the likely buyer is not a ChatGPT-style power user. Exa is selling to teams building internal agents, research products, data enrichment workflows and go-to-market systems that need repeatable output. The docs say Exa Agent can return natural-language answers or schema-validated JSON with grounded citations. Developers can pass an outputSchema
to force the response into a specific JSON shape, or use input.data
to bring their own rows for enrichment.
The examples in the launch post are pointed at budgets that already exist. Finance agents can retrieve fresh data from across the web and return it in a required format. GTM agents can enrich account or prospect lists, or generate lists of tens or hundreds of entities. The example Exa published asks the agent to research Databricks across funding, product launches, partnerships, executive hires, conference talks and public GitHub activity, then return a compact company brief with evidence.
Exa's bet is that owning retrieval will matter more than wrapping a model
Exa's strategy has been to own the search layer beneath agents rather than sit only at the application layer. In its Series C announcement on May 20, Exa said it raised $250 million at a $2.2 billion valuation led by a16z, and that Cursor, Cognition, HubSpot, OpenRouter, Monday.com and more than 400,000 developers were using Exa. The company also said more than 5,000 companies had used Exa since it announced a web search API for AI in early 2023.
Those adoption numbers are company-reported and not broken out for Exa Agent specifically. They do, however, explain the timing. A month after a large Series C, Exa is trying to move up the stack from raw search and content retrieval into the agent workflow where customers judge vendors on completeness, latency, citations and cost per finished task.
That move also puts Exa in a more direct fight with a crowded group of web intelligence providers. Parallel Web Systems sells a Task API for structured web search tasks and says its outputs include citations, rationale and confidence levels through its Basis framework. Perplexity offers Sonar for search-grounded generation. OpenAI exposes hosted web search through the Responses API, including controls such as domain filtering, sources and live access. Firecrawl is focused more directly on crawling and extracting web data for AI apps, and said in August 2025 that it raised a $14.5 million Series A led by Nexus Venture Partners.
Exa's differentiation claim is that it built an AI-native search engine and web-scale index rather than wrapping another search provider. In its Series C post, Bryk wrote that Exa's crawlers track more than 500 billion URLs and that the company trains specialized embedding models. Exa's about page similarly says the company tracks, indexes, crawls and monitors billions of webpages and serves them at high throughput.
The older origin story is useful because it separates Exa from companies that appeared after the agent boom. TechCrunch reported in 2024 that Bryk and Wang, best friends who met as Harvard freshmen, initially tried to build better search before AI companies started asking for API access. Bryk told TechCrunch the system was trained to predict the next link, not the next word, and Wang said the company launched before ChatGPT with a goal that was not originally to serve AIs.
The benchmarks are useful, but still self-published
Exa published benchmark comparisons against Perplexity Agent tiers, Parallel Task tiers and frontier models across WideSearch, BrowseComp, DeepSearchQA, FinanceAgent-V2, People-FindAll and Company-FindAll. The company says it uses row-level F1 on WideSearch, counting a row as correct only if the matched entity and all required enrichment columns are valid. Exa says the English WideSearch benchmark contains 100 tasks with required columns ranging from 3 to 14.
The methodology is stricter than cell-level scoring, which can reward scattered correct facts even when the agent fails to attach them to the right entity. But the results remain vendor-published. Parallel also publishes its own benchmark charts, including comparisons that put its processors ahead on certain tasks. The practical takeaway is not that one chart settles the market. It is that web research APIs are now competing on task-level economics, not just whether they can return a list of links.
Exa's April post on Highlights shows why cost is central to the pitch. Exa says its highlights model can reduce token usage by roughly 94% on some evals, and that 500 characters of highlights can match the accuracy of the first 8,000 characters of a page on SimpleQA. If that holds in production workloads, it gives Exa a way to sell better unit economics to agent builders who are already paying for model calls, search calls and downstream verification.
The open question is revenue quality. Exa has disclosed funding, valuation, developer usage and company usage. It has not disclosed ARR, gross margin, Exa Agent adoption, or how much of its usage comes from free-tier experimentation rather than durable production workloads. That is the next test for Bryk and Wang: whether Exa Agent becomes the default research substrate inside products that need the web, or one more tool developers benchmark, admire and swap out when the economics change.