Coding agents can miss repository changes made after their model's knowledge cutoff, so they may approve a dependency change that actually requires a migration.
2
A search tool alone does not make an agent search. The agent needs explicit rules that connect situations such as dependency bumps to a search action.
3
Exa returns query-dependent excerpts from large pages, giving the model about 500 characters instead of an entire 100,000-character page without adding latency to the search call.
Summary
Jakub Hojsan explains why coding and code-review agents need access to current information. Models can be released months after their knowledge cutoff, leaving them unable to review recent pull requests accurately. He walks through a dependency change that looks like cleanup to a model but actually requires a migration described in the repository's change log. Hojsan argues that adding web search is insufficient because the agent may not know when to use it. The agent needs rules, such as searching upstream whenever a diff changes a dependency or version. Exa's search returns the source trace, queries, sources, and selected highlights. Its query-dependent highlighting extracts the relevant lines from a large page through computation rather than another language model, so a model receives a small passage instead of an entire page. Hojsan also introduces Exa Agent, Exa's MCP access, and structured output features.
Model release dates leave code agents unable to review recent changes
Hojsan says models typically lag about six months between their knowledge cutoff and release date. That creates a gap when a repository receives a new change after the cutoff. He uses a recent PR from the Cooter vector store as an example that falls inside the blind spot after GPT-5.5's cutoff. During those months, real repositories keep changing. A model may be unable to create a new repository with the current information, and it may also be unable to review a pull request that depends on those changes.
A dependency change can look harmless while requiring a migration
Hojsan shows a diff that removes a parameter called inertia check. A human reviewer might search Stack Overflow for the error or check the upstream repository's change log. A model without web access may see a consistent-looking diff and call it cleanup. The actual reason for removing the parameter is a dependency bump that requires a refactor. Searching the repository and its breaking changes lets the agent explain the migration instead of accepting the change at face value.
Query-dependent excerpts give the model only the relevant part of a page
Exa does not return an entire GitHub page to the model. An interpretation step distills the page into the passage needed to answer the query. Hojsan contrasts roughly 500 characters with a 100,000-character page. Later, he says the query-dependent highlights are selected at runtime without using another language model. The same page can produce different excerpts for different requests, such as a biography query that returns hobbies or a query asking for a phone number.
A search tool needs rules that tell the agent when to use it
Hojsan says an agent can have web search access and still fail to search. He gives the example of asking Claude Code to use a model version that exists while the model says it does not. His proposed second step is to instruct the agent about search conditions. For code review, a rule could say that when a diff bumps a dependency or version, the agent should inspect the upstream source and ground its review in what it finds. The harness needs both the tool and the instruction to call it.
Exa exposes the search trace and keeps the API independent of model providers
Hojsan describes four benefits Exa gives an agent. It exposes the queries, sources, highlights, and full trace, unlike a native provider search that may return sources without the underlying content. It provides token-efficient excerpts. He also says its pricing can be more effective for sufficiently large workloads, and that a third-party API lets customers change between model providers while keeping one search interface and a fixed set of flexible parameters.
The same page can answer different queries without extra search latency
Hojsan demonstrates query-dependent highlights with one website. A biography query returns information about photography and motorcycle riding, while a phone-number query retrieves a different passage from the same site. He says the selection is computational and adds zero extra latency to the search call. This approach also applies to large code repositories, where passing every line to the model would waste context, and to news queries that require a particular person to be mentioned.
Exa Agent packages search across web and specialist data sources
Hojsan says coding is only one application. Cursor, Cognition, Warp, and Code Rabbit use Exa for web search, while Exa Agent is aimed at people who need a search experience without orchestrating search themselves. Exa Agent can combine web results with highlights from Similarweb for web analytics, Particle for podcast intelligence, and Crunchbase for private markets. Hojsan demonstrates a query for people who have mentioned the AI Engineer conference at its location, and says the approach can also find company employees and their contact information.
Exa MCP provides a direct way to add the search service to an agent
Hojsan says developers can call Exa through its MCP server and use it with providers that support MCP. He presents this as the way to start using Exa without building the full search orchestration themselves. In the questions, he explains that Exa turns a query into an embedding, combines semantic search with keyword filtering, and uses reranking stages. Its index is curated and contains tens of billions of documents, though he says it has fewer documents than Google.
Natural-language schemas can shape structured search results
In the final question, Hojsan describes a feature that generates a schema from a natural-language request. He says deep search supports up to 10 fields and agents support up to 100. After generating and confirming the schema, the system follows it. Users can also enable additional properties when they want more fields. His example mentions extracting hobbies, showing how the output format can be specified without manually writing the schema.
"The core problem that we have with large language models in code review and coding agent applications is that there's a knowledge cut off to all of these models."00:31
Who should watch
You are building a coding or code-review agent that can access search but does not reliably decide when to use it.
Your reviews involve dependencies, model versions, APIs, or repositories that may have changed after the model's knowledge cutoff.
You need a way to pass small, query-specific excerpts into a model and inspect which sources and passages produced its answer.
