Agents stop paging through files: Drex for Search picks the spot first, cutting steps 62% and input tokens 56% across 529 questions
Synopsis
When an AI agent needs an answer from a real folder of reports, ledgers and contracts, it searches, opens a file, reads pages and searches again, sometimes for twenty or thirty steps. Drex for Search moves that decision earlier: before the agent reads anything, the decision model Drex decides which folder, file and page can hold the answer, and inside a spreadsheet which table and column, so the agent reads only those places and returns the source. Across 6 professional-work benchmarks and 529 questions, coding agents using it took 62% fewer steps and 56% fewer input tokens, raised accuracy from 63.1% to 67.1%, and cut median time per question from 80.8 s to 35.5 s.
Interpretation
Agents now locate before they read, and answers come back with the file and page they came from. Keyword search such as grep or BM25 finds the word but not the meaning, while semantic search and RAG find similar-sounding text without knowing which document is right and cut tables into pieces, so the agent keeps running one more search, opening one more file, reading one more page. Across 6 professional-work benchmarks and 529 questions, the same coding agents with Drex for Search took 62% fewer steps and 56% fewer input tokens, raised accuracy from 63.1% to 67.1% (+4.0 points), and cut median time per question from 80.8 s to 35.5 s.
Locating happens in three steps: classify the question, narrow the workspace level by level, then answer from only the pages and tables that survive. The decision model Drex drives this: classification separates a quick lookup from a question that needs an agent to read documents and query tables, and narrowing keeps, at each level from folder to file to section to page and to table and column inside a spreadsheet, only what can hold the answer. The number of levels is fixed, so a million files take the same number of steps as a thousand; on OfficeQA Pro, 133 questions over 697 scanned PDFs, all ten agent and model pairs took fewer turns and fewer input tokens, eight answered more questions correctly, and average accuracy went from 59.3% to 61.9%.
The tool ships as a CLI, Python and TypeScript SDKs, an MCP server and a REST API, and works with Claude Code, Codex, Cursor, OpenCode and any agent that can run a command or call an MCP tool. Integration reuses the tool-calling path agents already have; the OfficeQA Pro agents reached their files through the Nace CLI. Each workspace has its own index built only from its own files, content never trains, tunes or evaluates shared models, and the tool can run in the customer's own cloud or data center when data cannot leave the network.
Perspective
The demonstrated setting is 6 professional-work benchmarks with 529 questions and ten agent and model pairs on OfficeQA Pro's 133 questions over 697 scanned PDFs, aimed at researchers and engineers who need agents to find answers in their own files.
The authors state that the Drex for Search results are one run while the as-shipped results average two runs, so accuracy should be read as ±2 points; input tokens count only the agent model's own, cache reads included, and Drex's own decisions are not counted. Two of the ten OfficeQA Pro pairs lost accuracy: Claude Code · Sonnet 5.5 fell from 69.2% to 66.9% and Codex · GPT-6 luna from 54.9% to 53.4%.