62% of AI Agent Search Steps Could Have Been Skipped

Nace AI has released Drex for Search, a retrieval tool for agentic search. The core idea: search quality is determined before the search begins. Drex makes decisions first, then lets the agent do the work.

Agentic search means an AI agent decides for itself what to look up, how to look it up, and when to stop, then synthesizes an answer. In reality, most agents rummage aimlessly through piles of files. One developer put it bluntly: poor retrieval is a common reason agents get questions wrong.

Drex's process has three steps:

1. Classify. Read the question and determine whether it is a quick lookup or a deep retrieval that requires reading documents and checking tables.
2. Narrow. Go layer by layer through folders, files, sections, and pages; for tables, go down to tables and columns. At each layer, keep only the content that may contain the answer. Depth is fixed, so a million files and a thousand files take the same number of steps.
3. Answer. Only the retained pages and tables go into the LLM, and answers come with file and page numbers.

Officially tested on 8 real professional-work benchmarks, leading on 6. On the 133 questions in OfficeQA Pro, the company says that 8 of 10 agent and model combinations answered more questions correctly, with steps reduced by an average of 62%, token consumption reduced by 56%, and accuracy increased by an average of 4 points. Gemini CLI rose from 54.9% to 61.7%, and steps dropped from 22.9 to 10.4. The only worsening entry in the table was Codex with GPT-6 luna, down 1.5 points.

Reading fast is not the key; the real difference is deciding what not to read. The problem with traditional RAG is that top-k never knows when there is enough information. Drex's stopping condition is simple: stop when you can cite a source. Someone in the brokerage business said that facts given by an agent must point to a specific email or document, or no one will trust them. Drex's citable stopping condition fits this need exactly.

Using it is not complicated. It supports Claude Code, Codex, OpenCode, Hermes, and others, via CLI, SDK, MCP, and REST. Upload files to a workspace, and the backend OCRs every page and builds a knowledge graph so terms interconnect across files. Ask a question, and the answer points to the file, page, line, or cell. It supports PDFs, scans, tables, audio and video, and more. Files stay in the workspace; the company says customer data is not used for training, and it has SOC 2 Type II.

A netizen summarized this logic in one sentence: a cheap decision is better than expensive wandering. Applied to agent search, this is more accurate than most architecture discussions.