Nace AI says coding agents using its new Drex for Search tool took 62% fewer steps and used 56% fewer input tokens to answer questions from business files, in its own tests across 529 questions, according to the company’s blog post of 26 October 2026. Nace launched the tool that day; TestingCatalog reported the launch.
Drex for Search is a search tool for AI agents. Nace says its decision model, Drex, picks the folder, file and page where an answer can be before the agent reads anything, and the agent answers with the file and page it came from.
Short video explainer
A vertical news Short narrated from the published headlines of this piece and three related FourWeekMBA pieces.
Business Pill · Context engineering
An evergreen Business Engineer explainer on context engineering, the general idea behind choosing what an AI model reads. It does not discuss Nace.
Drex for Search in three numbers, per Nace. Source: Nace AI blog, 26 Oct 2026 (company tests).
The key insight: As we read it, Nace is selling fewer tokens per answer rather than a better model: the company says Drex decides “which folder, which file, which page” before the agent reads anything, and the agent reads only those few places.
Nace says it compared coding agents as shipped with the same agents using Drex for Search, across six sets of professional work: Databricks’ OfficeQA, annual and quarterly reports, audit documents, ledgers, invoices, and accounts receivable and payable files.
Across all 529 questions, Nace says, accuracy went from 63.1% to 67.1%, and median time per question fell from 80.8 seconds to 35.5 seconds. Nace describes the accuracy change as “6% more answers right”; in percentage points it is 4.0.
The tool works with Claude Code, Codex, Cursor, OpenCode and “any agent that can run a command or call an MCP tool,” Nace says. It connects to AWS S3, Azure and DigitalOcean, according to TestingCatalog.
OfficeQA Pro accuracy for six of the ten agent-and-model setups Nace tested, as shipped and with Drex for Search, from Nace AI’s blog of 26 October 2026. Company tests; Nace says to read accuracy as plus or minus 2 points.
What Was Posted on X
Nace AI launched Drex for Search, a tool that lets agents retrieve answers from files stored in the cloud by giving them the exact pages they need.
When given a question, Drex for Search first finds the likely folder, then the file, then the page, and returns the answer, the quote, and the page number where it found the info.
> Drex Search is able to connect to a userโs AWS S3, Azure, and DigitalOcean, with the ability to read scanned PDFs, spreadsheets, email, and audio.
> Coding agents such as Claude Code, Codex, and Hermes Agent can use it as a tool, with Nace reporting 62% fewer steps and 56% fewer input tokens per question.
> Across 529 queries from six professional work benchmarks, including Databricksโ OfficeQA, agents using it answered 67.1% correctly, up from 63.1%.
— ๐จ AI News | TestingCatalog (@testingcatalog) Oct 26 15:10
TestingCatalog reported that Nace AI launched Drex for Search, a tool that lets agents retrieve answers from files stored in the cloud, connecting to AWS S3, Azure and DigitalOcean, with Nace reporting 62% fewer steps and 56% fewer input tokens per question.
Agent by Agent
On OfficeQA Pro, Databricks’ benchmark of 133 questions over 697 scanned PDFs, Nace says it ran ten combinations of coding agent and model. It says every one took fewer turns and used fewer input tokens, and eight of ten answered more questions correctly.
Nace’s own table shows two that did worse: Claude Code with Sonnet 5.5 went from 69.2% to 66.9%, and Codex with GPT-6 luna from 54.9% to 53.4%. Claude Code with Opus 5.5 went from 66.5% to 69.9%, with turns per question falling from 18.5 to 9.1.
Nace adds its own caveat: the as-shipped results are the average of two runs on 20–22 October and the Drex results one run on 24–25 October, “so read accuracy as ±2 points.”
Who Nace Is
Nace’s blog lists a $21.5 million funding round announced in May 2026, an earlier decision model called Drex in September and a multimodal Drex Max on 19 October. The company sells the search tool through a command-line tool, SDKs, an MCP server and a REST API.
As we read it, the headline numbers are averages across agents: Nace says across all 529 questions accuracy went from 63.1% to 67.1%, while its own OfficeQA Pro table shows two of ten setups scoring lower.
As we read it, the cost case rests on tokens: Nace says agents used 56% fewer input tokens, and on OfficeQA Pro Claude Code with Opus 5.5 went from 906k to 475k input tokens per question.
As we read it, the comparison has a timing caveat Nace states itself: the as-shipped runs are the average of two runs on October 20–22 and the Drex runs one run on October 24–25.
Nace AI blog, 26 October 2026
“Drex decides where the answer can be”
Three Implications
FOR TEAMS RUNNING CODING AGENTS ON DOCUMENTS Nace says Drex for Search works with Claude Code, Codex, Cursor, OpenCode and agents that can call an MCP tool.
FOR FINANCE AND AUDIT FILES Nace says its benchmarks include audit working documents, ledgers and trial balances, invoices, and receivable and payable files.
FOR AGENT COSTS Nace says median time per question fell from 80.8 seconds to 35.5 seconds across the three agents in its timing table.
The framework puts it this way: “The new job of finance is to manage the ratio between the two — work produced per token burned.”
As we read it, Nace’s pitch targets that ratio: the company says its tool cut input tokens per question by 56%. The framework was published in June 2026 and does not discuss Nace.
What Is Not Established
Every number here is Nace’s own, from its blog post; we did not see an independent test. The accuracy comparison mixes runs from different days, which Nace flags itself. Nace gives no pricing in the post, and no customer names.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Nace AI says its Drex for Search tool cut coding agents’ steps by 62% and input tokens by 56% across 529 business-document questions, with accuracy rising from 63.1% to 67.1%, according to its blog post of 26 October 2026.
A note on sourcing. Every figure here is from Nace AI’s own blog post of 26 October 2026, from tests the company ran itself; the cloud connectors are from TestingCatalog’s post. This is not investment advice. We did not contact Nace AI.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
