- A speaker from Exa argued that agents need a few hundred characters chosen for the question, not a whole page or a list of links.
- Adding a search tool is not enough: the speaker said agents also need explicit rules for when to search.
Search built for AI agents may hand back a small piece of a page, chosen by the question, instead of ten links for a person to click. A speaker from Exa, a search company, made this case on the AI Engineer show. The talk was about coding and code-review agents.
What was said
The speaker described himself as a forward deployed engineer at Exa. He argued that coding agents "don't really need these 10 blue links". Exa instead takes a query and returns highlights to pass into a language model. An interpretation step shrinks the page to what the model needs. He said the model no longer receives a 100,000-character page. By his account it gets "quite literally only 500 characters" instead.
The speaker said this step is computational. It pulls certain lines at runtime without using a language model, so it adds no extra delay. The snippet depends on the question. On one personal web page, a biography query surfaced hobbies. A query for a phone number surfaced different text from the same page.
His main example was a code review. A model without web search saw a removed parameter and called it a cleanup. The real cause was a dependency upgrade the model had never seen. He added that bolting on a search tool is not enough. Agents need rules, such as checking the upstream source when a change bumps a version.
Why it matters
Our reading: if the reader is a model, the question for a site owner may shift from whether a page ranks to whether the right passage is easy to pull out. Documentation teams may need clear, self-contained passages. For teams building agents, the talk's point is that search is a design decision, not a checkbox. Someone has to decide when the agent searches.
Buyers have a transparency question too. The speaker said built-in search tools from big model providers act like a black box. They can take up to 10 seconds and return some sources but none of the content. He said Exa gives a full trace of queries, sources and highlights, which can feed a team's own monitoring.
The other side
The speaker works for the company selling this. The talk gave no independent test of the zero-added-delay claim or of answer quality. The 500-character figure came from one example, not an average across queries.
Questions also went unaddressed. If software picks the snippet, a wrong cut can mislead the agent, and the talk did not say how often that happens. It also did not cover what happens to the sites whose pages are cut down, or how their owners are paid or credited.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Your Coding Agent Is 6 Months Out of Date — Jakub Hojsan, Exa (2026-10-02)
