Patches, Updates & Engineering Notes
Release notes and engineering write-ups from AgentDoc – the AI-native docs editor for LLM agents and voice-first users (also known as agent doc, agentdocs, docedit). We ship in small, frequent patches; here is what changed and why.
DOCX Import: Round-Tripping Word Documents Through an AI Editor
The mirror of the export rewrite. Page breaks, hyperlinks, headers/footers, fonts, colours and line spacing all preserved on the way from Word into the editor — and back out again. Plus the iterating-XML-children trick that stops python-docx from silently dropping every hyperlink.
Rebuilding PDF + DOCX Export: WeasyPrint, Pandoc Lua Filters, and Trusting the Frontend
We rewrote the export pipeline so PDF and DOCX downloads actually match what users
see on screen. The architectural insight: the browser's paginator is already the
source of truth – the server just needs to follow it. Plus a tour of the Pandoc
reference.docx trick and a Lua filter that translates the editor's
stand-off classes into native Word run properties.
AgentDoc Patch Notes – April 2026: Renderer Fixes, Decoration Composition, Toggle Semantics
Five fixes shipped this week: nested-list depth preservation, decoration-* class composition,
proper TOGGLE semantics on apply_* tools, inline-code escaping, and decoration styling
generalised beyond <span>. What broke, why it broke, and what we changed.
Tool Granularity in LLM Agents: What 20 Workflow Configurations Taught Us
We benchmarked atomic tools vs. composite macros vs. State-Constrained ReAct FSMs across 20 workflow configurations (A–T) and 13 scenarios. The result is not what the "fewer tools = better" narrative predicts. Detailed write-up of accuracy, token cost, and where the cliffs are.
Voice-First Document Editing with Gemini Live: Architecture & Tradeoffs
Why we wired Gemini 3.1 Flash Live native audio directly to the MCP server instead of stacking a second LLM behind a transcription proxy – and what that buys you in latency, accuracy, and tool-call reliability.
Want updates? Bookmark this page – we publish about once every 2–3 weeks. For machine-readable context see /llms.txt.