Jo Kristian Bergum defines agentic search as retrieval inside an agent's task loop, with the model, harness and engine each contributing to success. BM25's scoring function has not changed, but agents can write longer, more precise queries and repeatedly refine them.
Jo Kristian Bergum uses BrowseComp+ to distinguish reasoning with supplied evidence from finding that evidence through a search tool. He explains why document length and BM25 parameters matter, and argues that exact matching remains useful for names and identifiers. This is not a claim that every BM25 setup outperforms embeddings.
Jo Kristian Bergum describes retrieving documents into a file-system workspace, then using titles, snippets and familiar text tools to inspect selected material. He favors judging the entire search trajectory by task success. Hornet's serving-efficiency comparison is company-presented evidence, not an independently verified universal result.
Watch on YouTube
