TriAgent: deployed multi-agent system triages support tickets to find and fix stale documentation
Researchers deployed a multi-agent system that mines support tickets for signs of outdated documentation, then drafts fixes for human review.
In plain words
Software teams often update code faster than the documentation that explains it, and AI assistants that read that documentation can then repeat outdated information. The described system, TriAgent, automatically scans support tickets to spot where documentation has fallen behind and proposes repairs. A human still reviews each proposed fix before it is applied.
What was announced
A paper on arXiv describes a deployed multi-agent system built around a triage agent called TriAgent, which continuously extracts signals of documentation drift and repairs it. The authors argue that as users rely on LLM-based agents that treat documentation as their primary interface, stale information propagates and is treated as authoritative. Signals include explicit ones, such as resolved support tickets, and implicit ones, such as code-documentation discrepancies detectable through automated loops.
In 43 days of shadow-mode operation, the system triaged 499 tickets and surfaced 92 documentation gaps. A panel of LLM judges from independent model families rated 81.1% of TriAgent's verifiable assertions as correct, and on length-comparable pairs preferred its responses over human engineers' replies in 70.5% of cases. The authors note the two were often complementary, with humans surfacing insights missing from existing documentation. After a repair, blinded judges preferred the agent's post-repair answer to its own pre-repair answer 71.4% of the time (n=35), which the authors describe as evidence of self-improvement confirmed by a blinded human audit.
What it does not show
- The results come from a single arXiv abstract; the full paper's methods, deployment context and evaluation details are not available in the supplied excerpt.
- The judges were LLM-based rather than independent human raters for the main metrics, and the abstract does not report how the 92 surfaced gaps were resolved after human review.
- Shadow-mode operation means the system ran alongside existing processes, so effects on live documentation workflows are not established.
For security and IT teams
- Scope
- The reported deployment ran in shadow mode for 43 days, triaging 499 tickets and surfacing 92 documentation gaps; it is described in an abstract only.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.