Jo Kristian Bergum

Bergum’s first-person career account describes starting in enterprise search at Fast Search & Transfer and leaving in 2006. His later work includes Yahoo and the chief-scientist role documented by Vespa’s original author pages. At Hornet, he builds for agents that issue long, structured queries and may perform many searches inside one task.[3][5][4][2]

Let code manage the evidence before it enters context

In the essay coauthored with Skip Everling, an agent writes a retrieval loop that can issue queries, deduplicate candidates, inspect fields and preserve only selected evidence. Ranking and freshness remain programmable controls inside that loop. Keeping bulky intermediate results outside the model’s conversation reduces the amount it must repeatedly read, while structured results let it decide what to inspect next. The authors emphasize that a better interface still depends on the underlying engine returning relevant documents.[6]

Hold the search engine fixed and change the harness

Bergum’s follow-up compares one tool call per model turn with a bounded Python program that can perform several retrieval steps. Both use the same model, corpus, lexical retriever and session limits. On the first one hundred BrowseComp-Plus questions, the code interface increased the share of labeled documents placed in the final curated set while reducing prompt tokens and cost. That metric measures retrieval recall, not the correctness of a final answer, and the experiment covers one model and retrieval configuration.[7]