Mind the Gap: A Deployed Multi-Agent System for Documentation Maintenance

According to the abstract, Lorenzo Masoero describes a multi-agent system that finds and repairs out-of-date software documentation. Over 43 days in shadow mode it triaged 499 tickets and surfaced 92 gaps, with LLM judges rating 81.1 percent of its checkable assertions correct.

Why it matters

The same thing, for different readers.

What it means for you

Stale help pages mislead both people and the AI assistants that read them, and this work reports an automated way to spot where documentation has fallen behind.

For researchers

The reported 81.1 percent accuracy comes from a panel of LLM judges rather than human review, so the correctness figure rests on model grading.

How the finding was reached

In our words, from the source. Weigh the finding by how it was made.

What was done
A multi-agent system built around a triage agent, run in shadow mode for 43 days on support tickets and code-documentation discrepancies; assertions judged by a panel of LLM judges from independent model families.
What it found
It triaged 499 tickets and surfaced 92 gaps; 81.1 percent of verifiable assertions (245 of 302) were judged correct by a majority of three judges, as the author reports.
What it does not show
Read from the abstract excerpt only; shadow mode means outputs were not applied, and grading used LLM judges.
Status
A preprint: not yet peer reviewed

The record

Questions
What happens to work, income and the economy?
People
Lorenzo Masoero (Wrote it)
What we read
We have read only the description the publisher gives. No quotations until someone reads the source itself.
Found by
The Radar, run of November 18, 2026, via arxiv-ai-safety-query (tier 1: the source’s own channel or an official record)
Cite this
arXiv, “Mind the Gap: A Deployed Multi-Agent System for Documentation Maintenance”, November 17, 2026. Entry in the Ctrl AI Base, https://ctrlai.com/base/arxiv-2026-11-17-multi-agent-documentation-maintenance
Our notesLight lane: written from the triage excerpt only; nothing fetched and no source text saved.