Agentic search is an approach to retrieval in which an agent gathers its own context by searching a
file system with ordinary tools, deciding for itself what to pull into its context window. In
Anthropic’s description, the file system represents information that could be pulled into the
model’s context: when Claude encounters large files such as logs or user-uploaded files, it decides how
to load them by using bash tools like grep and tail. On that account the folder and file structure
an agent works in becomes a form of Context engineering in its own right.
Usage
Anthropic uses the term in Building agents with the Claude Agent SDK as one of the ways
an agent built on the Claude Agent SDK gathers context, the first stage of the
gather-context, take-action, verify-work loop the post describes. Its worked example is an email agent
that stores previous conversations in a folder named Conversations, so that it can search them when
asked about earlier exchanges.
The term is defined there by contrast with semantic search, which chunks the relevant context, embeds the chunks as vectors and searches for concepts by querying those vectors. The post characterises semantic search as usually faster than agentic search but less accurate, more difficult to maintain and less transparent.
When It Applies
Anthropic’s recommendation is to start with agentic search and to add semantic search only if faster results or more variations are needed. The approach assumes an agent that has a file system and shell tools to search it with, and it trades speed for the accuracy, maintainability and transparency the post attributes to it. How well-established it is: the comparison with semantic search is stated as Anthropic’s own guidance, drawn from its teams’ experience building agents on the SDK, and the post gives no measurements for it.