The software, not the model, decides whether an AI agent wrecks your Excel: an open source fork proves it

By Zendoric · September 23, 2026.

Paper Instruments, a startup little known until now, has released Paper Office: three open-source Python packages —paper-docx, paper-pptx and paper-xlsx— that extend the standard python-docx, python-pptx and OpenPyxl libraries so AI agents can edit Word, PowerPoint and Excel documents without wrecking them internally. According to the company's own report, across five models and 61 editing tasks on existing files, Paper Office raised the success rate to 92.5% of attempts, against 80.7% for the original libraries unaided and 69.5% with Anthropic's official Office skills kit.

That last figure is the most striking: Anthropic's official add-on for working with Office files performed worse than using no add-on at all. It should be read with caution — it is a number published by a company that competes directly in this space with its own product — but the underlying pattern Paper Instruments describes is consistent and technically verifiable: when the library does not know how to do something, the model improvises, and that improvisation comes at a cost.

.docx, .pptx and .xlsx files are not plain text: they are ZIP archives containing multiple XML fragments linked to one another (the Office Open XML standard, or OOXML). Changing a single visible word can require touching several pieces at once — the link between a comment and its reply thread, a chart's cache, a pivot table's range. The standard libraries, with no major updates in years, do not cover many of those cases. When the agent runs into an operation the library does not support, it falls back on writing code that edits the file's internal ZIP directly — and that, according to Paper Instruments, is where most of the silent errors creep in: a comment thread that loses its link, a formula that stops recalculating, a chart still pointing at old data.

The figure that illustrates the mechanism: agents resorted to editing the file's internal ZIP in 78.7% of attempts with no helper package, in 50.5% with Anthropic's skills, and in only 1.6% with Paper Office. The improvement did not come from smarter models — Opus 5, GLM 5.3 Flash, GPT 5.6 Sol, Grok 4.6 and DeepSeek v4.1 Flash were tested, with gains in 12 of the 15 model-and-format combinations — but from software that handles the mechanical, risky part of editing on the agent's behalf: locating the right object, applying the change, and checking that everything else remains intact after saving.

Paper Instruments frames the problem as the reason AI agents still barely penetrate the daily work of consultants, lawyers, bankers and traders: they do not edit existing documents the way a person would, and the result —a slide, a spreadsheet, a contract— lands in an uncanny valley that is no good for showing a client. It is what the company calls "brownfield knowledge work": not creating documents from scratch, but touching those that already exist, with their history, their inherited templates and their dependencies, without leaving any trace of error.

Here is the reading that interests us. In recent months we have watched the raw quality of frontier models converge; they now compete within a narrow margin on the benchmarks that actually measure something. When the model stops being the variable that discriminates most, the battle moves one floor down: to the plumbing. We saw it with Google's and Microsoft's repositioning around agent integration standards, and we see it again here, in miniature: whoever builds the software that keeps an agent from wrecking a real Excel file gains more practical ground than whoever simply trains a slightly better model.

It also confirms something we argued in earlier editions: in domains where getting it wrong is costly — a badly edited contract, a spreadsheet that reaches a board of directors with an out-of-sync cell — the product answer is not "more intelligence," it is defensive architecture: validate before touching, preserve what you were not asked to change, and explicitly refuse when the operation cannot be done with guarantees. Paper Office institutionalizes this with edit receipts, package diffs and a "certification" layer that recalculates formulas with LibreOffice to confirm the result adds up. It is unglamorous engineering, but it is exactly the kind of work that separates a demo from a product a law firm or an investment bank dares to put in an agent's hands without constant supervision.

The finding is worth keeping in proportion: it is a study by a small startup about its own product —one that also develops an agent harness, Feather, where Paper Office comes integrated by default— with 61 tasks and no independent review as yet. It is not proof that the problem is solved, and it remains to be seen whether other developers adopt it or replicate it with their own data. But the underlying trend —that low-level software, not just the model, determines whether an agent can work reliably on real documents— will probably become widespread, with or without Paper Instruments.

If we really want agents to free professionals from mechanical work so they can devote themselves to judgment and client relationships —the underlying promise of this transition— they first need to stop breaking what they touch. That is an engineering requirement before it is a matter of intelligence, and this week someone has put it in writing, with numbers.