Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa

Every model has a knowledge cutoff, and according to Jakub Hojsan, models reach users roughly half a year after it. A code review agent built on one of them is often judging changes that are newer than anything it learned. Hojsan, a forward deployed engineer at Exa, spoke about this at the AI Engineer World's Fair 2026 in San Francisco, and the talk explains how Exa approaches the problem with search designed for agents.

The opening section makes the gap concrete. From about 1:12 he takes an actual pull request that sits outside the model's knowledge, then contrasts how a person would check it with the reason the model gets the review wrong.

The obvious fix is to give the agent a web search tool. Hojsan says that by itself does not solve anything: the agent also needs written guidance about which situations call for a lookup. That guidance is the subject of the chapter at 3:56.

His second point concerns what the model reads once it has searched. Rather than dropping an entire web page into the context, Exa returns short highlights chosen for the specific query. The description puts this at a 100,000-character page cut to around 500 characters without extra latency. That figure is Exa's own, and the talk includes a demo of the feature.

Hojsan also compares Exa with the web search that model providers build in, on transparency, cost and not being tied to one model vendor. Later chapters go beyond coding with a preview of Exa Agent and a demo that hunts for conference speakers. Connecting through the Exa MCP server comes up at 9:25. Two audience questions close the session, one on ranking billions of documents and one on structured output schemas.

The talk runs about 12 minutes. It suits developers building coding or review agents that stumble on recent changes, and anyone deciding when an agent should search and how much text it should get back. Pricing is not covered in the sources.