Paper Instruments teaches agents to edit Office files without dismantling them

Why it matters

Paper Instruments is betting that agents will need purpose-built file libraries, not just better prompts, to make safe edits to the Office documents professionals already depend on. Its reported benchmark supports the case, but combines software with guidance, leaving real-world reliability and developer adoption as the proof points.

On September 23rd, Paper Instruments released Paper Office, three Python packages designed to help AI agents edit existing Word, PowerPoint and Excel files while preserving the parts users did not ask them to change. The release puts co-founder and CEO Daanish Khazi, Gavin Bains and Joseph Besgen behind a practical bet: agents need better software for handling the documents professionals already use, not just better instructions for generating new ones.

Khazi and his co-founders came through Y Combinator's Spring 2025 batch under the name The LLM Data Company. YC describes that earlier work as tooling for writing, versioning and evaluating models and agents. Paper Instruments now works on open-source models and tooling for knowledge work. Paper Office carries that interest in measuring and improving agent performance into a lower layer of the stack: the Python libraries an agent uses to make an edit.

The hard part is keeping what was already there

A DOCX, PPTX or XLSX file is a package of connected XML files, images and other resources, rather than a single document body. A seemingly simple edit can touch formatting, comments, formulas, charts or relationships among those components. Paper Instruments' argument is that agents often reach for raw file internals or lossy conversions when their tools cannot express the requested change safely.

The three packages extend familiar Python libraries: paper-docx builds on python-docx, paper-pptx on python-pptx, and paper-xlsx on openpyxl. They keep the usual import patterns while adding ways to inspect hidden structure, target edits, preserve package contents, report changes and refuse operations the software cannot handle safely, according to the announcement.

That refusal behavior is central to the pitch. In a spreadsheet, an edit that changes visible cells but breaks a formula reference can be worse than no edit. In Word, replacing text that spans multiple runs can disrupt formatting or review comments. Paper Instruments says its packages expose structure to the agent and validate supported edits before saving. For Word, features include cross-run search, tracked edits and comment threads. Its Excel package adds reference-aware row and column changes and package-preserving saves; the PowerPoint package supports formatting inspection and relationship-safe slide operations.

The product is aimed at developers building agents for consulting, legal, banking and operational work, where the source file is often a template or live working document rather than a blank page. That focus gives Paper Instruments a specific wedge: help agents handle the messy files already in circulation without requiring users to accept a converted approximation.

A benchmark with a bundled intervention

Paper Instruments reports that its packages, used with accompanying guidance, passed 92.5% of trials across five models and 61 tasks. It compares that result with 80.7% for upstream libraries without skills and 69.5% with Anthropic's comparable Office skills. The company also says agents directly edited Office file internals in 1.6% of Paper runs, against 78.7% without skills and 50.5% with Anthropic skills.

Those numbers describe a combined setup: the package plus guidance. They do not isolate the library from the instructions around it, so the reported results are best read as a comparison of agent tool configurations. Paper Instruments' post does not name the five models in the excerpted results, and the headline percentages should not be mistaken for an independently established benchmark. The company frames the test as evidence that software itself can improve agent performance; the release offers a useful hypothesis for developers to test against their own documents and workflows.

The comparison also places Paper Office beside Anthropic's Agent Skills, which provide procedural instructions and resources for tasks such as creating spreadsheets and presentations. Paper Instruments is competing at a different layer: skills tell an agent how to approach a workflow, while these packages aim to give it safer operations on the files. In the company's framing, better prompts cannot by themselves preserve every relationship in a complex Office archive.

For Khazi and his co-founders, Paper Office also moves their work closer to the software that controls agent behavior. The earlier LLM Data Company focused on evaluation tooling; Paper Office supplies a specialized tool layer and measures how agents perform with it. The two efforts share an underlying premise: professional AI depends on the surrounding system as well as the model.

The test of the premise will be whether developers trust these packages with consequential documents. Passing a bounded task suite is one milestone; reliably preserving a client's templates, spreadsheet dependencies or tracked revisions across varied real-world files is another. Paper Instruments has released code developers can examine and use, and its explicit refusal paths make the safety claim concrete enough to inspect. The next proof will come from adoption and from how the packages behave on documents outside the authors' test set.