Paper Office's 92.5% agent pass rate measures its Office libraries and their guidance as one bundle

A DOCX, PPTX or XLSX file is a package of connected XML files, images and other resources ^[4]. A small edit can touch formatting, comments, formulas, charts or the relationships among those parts ^[4]. In Word, replacing text that spans several runs can disrupt formatting or review comments ^[9]. In Excel, an edit that updates the visible cells but breaks a formula reference can be worse than no edit ^[9]. Paper Instruments argues that when an agent's library cannot express a change safely, the agent falls back on raw file internals or a lossy conversion ^[5].

The design choice I like most is the refusal. According to the announcement, the packages validate supported edits before saving and refuse operations they cannot handle safely ^[7]. Each feature targets a place where its format breaks: cross-run search, tracked edits and comment threads in Word; reference-aware row and column changes and package-preserving saves in Excel; formatting inspection and relationship-safe slide operations in PowerPoint ^[8]. A refusal is an error the agent can read and act on. A saved workbook with a broken reference gets found later, by a person. Paper is aiming at developers building agents for consulting, legal, banking and operational work, where the file is often a template or a live working document ^[13].

Adoption looks cheap. The packages build on python-docx, python-pptx and openpyxl and keep their import patterns ^[6]. If that holds, trying one is close to a dependency change.

The benchmark is harder to read. The 92.5% comes from the packages used with accompanying guidance ^[10]. As RuntimeWire notes, the results do not isolate the library from the instructions around it ^[15]. Against upstream libraries with no skills, the bundle adds 11.8 points ^[1] and cuts the failure rate from 19.3% to 7.5% ^[4]. Against Anthropic's Office skills it leads by 23 points ^[3].

Upstream agents with no skills edited file internals directly in 78.7% of runs and still passed 80.7% of trials ^[10][11]. Adding Anthropic's skills cut raw edits to 50.5% of runs and dropped the pass rate to 69.5%, a loss of 11.2 points ^[11][2]. Anthropic's skills are procedural instructions for tasks such as creating spreadsheets and presentations ^[14]. Paper's tasks edit existing files ^[1]. The Anthropic result helps Paper's case, because guidance on its own gave no free lift here. It also shows guidance can move this score by double digits, so Paper's own guidance could account for part of its 11.8-point gain ^[1].

For the headline number to transfer to another team, their model has to behave like the five Paper tested. The excerpted results do not name those models ^[12]. Their templates have to resemble Paper's 61 tasks ^[10]. And the gain has to come from the code they would install. Five models across 61 tasks is 305 model-task combinations ^[6].

Two more arms would separate the code from the guidance: Paper's packages with no guidance, and the upstream libraries with Paper's guidance. The founders came through Y Combinator's Spring 2025 batch as The LLM Data Company, which YC describes as tooling for writing, versioning and evaluating models and agents ^[3]. Running an ablation is squarely their old line of work.