Drex for Search goes generally available: agents pick the page before reading, cutting steps 62% across 529 professional questions
Synopsis
Drex for Search is a search tool for AI agents in which Drex, a decision model, decides before the agent reads anything where the answer can be: which folder, file and page, and inside a spreadsheet which table and column. Across 6 professional-work benchmarks and 529 questions, coding agents using it raised accuracy from 63.1% to 67.1%, took 62% fewer steps, used 56% fewer input tokens, and cut median time per question from 80.8 s to 35.5 s. On the 133 questions of OfficeQA Pro, all ten agent-and-model pairs took fewer turns and fewer input tokens, and eight answered more questions correctly.
Interpretation
AI agent file search gains a prior decision layer: before reading any page, the tool determines which folder, file and page can hold the answer, and inside a spreadsheet which table and column, so the agent reads only those few places and returns the file and page the answer came from. Keyword search such as grep or BM25 finds the word but not the meaning, while semantic search and RAG find text that sounds similar but do not know which document is the right one and cut tables into pieces, so the agent keeps going with one more search, one more file, one more page. Across 6 professional-work benchmarks and 529 questions, coding agents as shipped versus the same agents with Drex for Search went from 63.1% to 67.1% accuracy, took 62% fewer steps and used 56% fewer input tokens.
The procedure runs in three steps: classify the question, squeeze the workspace level by level, then pass only the kept pages and tables to the LLM for the answer. Drex is the decision model behind the tool; it first reads the question and decides whether it is a quick lookup or one that needs an agent to read documents and query tables, then keeps only what can hold the answer at each level of folder, file, section, page, and inside a spreadsheet table and column. The number of levels is fixed, so a million files take the same number of steps as a thousand; on the 133 questions of OfficeQA Pro over 697 scanned PDFs, ten agent-and-model pairs cut input tokens per question by 57% and turns by 53%, with average accuracy rising 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. Each workspace has its own index built only from its own files, only what you want searched is uploaded, content never trains, tunes or evaluates shared models, and the tool can run in your own cloud or data center when data cannot leave the network. On public leaderboards it ranks first on MADQA at 94.7%, first on VisDocAgentBench at 77.5 R@1, second on BrowseComp-Plus at 92.05% with 16.6× fewer calls than the first place, and third on EnterpriseRAG-Bench at 88.22, as published on October 20, 2026.
Perspective
The demonstrated scope is 6 professional-work benchmarks totaling 529 questions, plus ten agent-and-model pairs on the 133 questions of OfficeQA Pro over 697 scanned PDFs, where agents could reach the Drex for Search workspace only through the Nace CLI and everything else stayed the same.
Accuracy should be read as ±2 points: as-shipped results average two runs on October 20–22, 2026, while Drex for Search results are one run on October 24–25; input tokens count only the agent model's own, cache reads included, and exclude Drex's own decisions. Two of the ten OfficeQA Pro pairs lost accuracy, with Claude Code · Sonnet 5.5 falling from 69.2% to 66.9% and Codex · GPT-6 luna from 54.9% to 53.4%.