Ryan Ahern of Exa introduces Exa Snapshot, which lets agents search the web as it looked on any past date. In a talk, he argues search quality matters more as agents broker most searches, and shows how teams use point-in-time results to build predictive signals and backtest them without overfitting.
A talk and hands-on walkthrough by Ryan Ahern of Exa showing how to backtest agent predictions against a past snapshot of the web. LLM agents predict events like company acquisitions from signals such as leadership turnover, evaluated against versioned search results. It solves look-ahead bias for developers tuning agents on historical data.
A talk and hands-on walkthrough by Ryan Ahern of Exa showing how to backtest agent predictions against a past snapshot of the web. LLM agents predict events like company acquisitions from signals such as leadership turnover, evaluated against versioned search results. It solves look-ahead bias for developers tuning agents on historical data.
A talk by Exa founder Will Bryk explains the company's pivot from a 2022 consumer search engine to an API for AI agents. Bryk argues 2026 is the year machine-issued searches exceed human searches, and that search must be rebuilt to answer queries rather than recommend results.
A talk by Exa founder Will Bryk explains the company's pivot from a 2022 consumer search engine to an API for AI agents. Bryk argues 2026 is the year machine-issued searches exceed human searches, and that search must be rebuilt to answer queries rather than recommend results.
In a talk recorded at AI Engineer World's Fair 2026, Exa forward deployed engineer Jakub Hojsan explains how models ship roughly six months behind their knowledge cutoff, leaving code-review agents unable to assess recent changes. He presents Exa's approach: agent rules for when to search, plus query-dependent highlights that compress a 100,000-character page to about 500 characters without added latency. The talk matters to practitioners building agents on top of models with stale training data, and previews Exa Agent.
In a talk recorded at AI Engineer World's Fair 2026, Exa forward deployed engineer Jakub Hojsan explains how models ship roughly six months behind their knowledge cutoff, leaving code-review agents unable to assess recent changes. He presents Exa's approach: agent rules for when to search, plus query-dependent highlights that compress a 100,000-character page to about 500 characters without added latency. The talk matters to practitioners building agents on top of models with stale training data, and previews Exa Agent.
Telem is a web search open router that routes agent web search queries across multiple providers (Exa, Parallel, Tavily, Brave, SerpAPI) and traces the results with quality metrics. It helps developers visualize whether a bad agent run is due to web search or reasoning, solving the problem of debugging agent performance. Built by a VC with an information retrieval background, it targets developers building LLM agents.
Telem is a web search open router that routes agent web search queries across multiple providers (Exa, Parallel, Tavily, Brave, SerpAPI) and traces the results with quality metrics. It helps developers visualize whether a bad agent run is due to web search or reasoning, solving the problem of debugging agent performance. Built by a VC with an information retrieval background, it targets developers building LLM agents.
An arXiv paper studies why search-augmented LLMs produce outdated answers even when relevant evidence is retrieved. The authors built temporal facts whose values changed after training cutoffs, evaluating one proprietary model with deployed web search and six open-weight models under controlled retrieval. They find models often answer from search snippets without opening pages, leading to outdated answers and unfaithful citations — a gap between evidence availability and its effective use.
An arXiv paper studies why search-augmented LLMs produce outdated answers even when relevant evidence is retrieved. The authors built temporal facts whose values changed after training cutoffs, evaluating one proprietary model with deployed web search and six open-weight models under controlled retrieval. They find models often answer from search snippets without opening pages, leading to outdated answers and unfaithful citations — a gap between evidence availability and its effective use.