Paper Office Python packages enable AI agents to manipulate Word, PowerPoint, Excel files

New Python Packages for Agent-Based Document Editing

Paper Office is a newly released suite of Python packages designed to enable AI agents to manipulate Microsoft Word, PowerPoint, and Excel files. These packages build upon existing open-source libraries such as python-docx, python-pptx, and OpenPyxl, aiming to enhance safety, correctness, and breadth in agent-driven document editing tasks.

Performance Improvements and Reduced Code Generation

Across five models and 61 tasks, Paper Office packages, combined with guidance, achieved a 92.5% success rate in trials. This compares to 80.7% for upstream packages without specific skills and 69.5% for Anthropic's comparable Office skills. A key improvement is the reduction in agents writing direct code to edit Office file internals, occurring in only 1.6% of Paper runs, versus 78.7% without skills and 50.5% with Anthropic skills.

Addressing Challenges in Agent Document Manipulation

The development addresses the challenge of AI agents manipulating existing documents in a way that maintains fidelity to human workflows. Current agent methods often produce documents that are not suitable for professional use due to inconsistencies. Office Open XML (OOXML) files, which are ZIP archives of XML files and other resources, require careful manipulation to maintain consistency across linked components, even for minor visible changes.

Reworking Existing Libraries for Agent-First Use Cases

Standard Python Office libraries, while mature, have not seen updates in several years and lack feature breadth and correctness for many agent-specific workflows. Paper Office forked, patched, and reworked the APIs of these libraries to support agent-first use cases, focusing on improving correctness and expanding feature completeness. This allows agents to find logical objects, apply changes while maintaining formatting, and preserve unaffected content.