What the study found
The study reports that TSGuard, a user-centric multi-agent system, can diagnose incidents for AI workloads in the cloud immediately for users who deploy the workloads. It is described as outperforming current baselines in evaluation on Microsoft Azure incident records.
Why the authors say this matters
The authors say the current provider-centric incident workflow can take several days because troubleshooting is manual and many incidents must be handled. They suggest TSGuard matters because it gives users direct, immediate diagnosis and may reduce operational delays and productivity loss.
What the researchers tested
The researchers presented TSGuard, which uses two phases: an offline phase that mines historical on-call experiences to build domain-specific knowledge bases, and an online phase that mimics human expert diagnosis through structured reasoning and iterative trial-and-error. They evaluated it using production incident records from Microsoft Azure.
What worked and what didn't
TSGuard improved diagnostic accuracy by 19.8% compared with state-of-the-art baselines. It also reduced average verification time by 63.4% compared with the sequential execution baseline.
What to keep in mind
The available summary does not describe detailed limitations beyond the evaluation setting. The reported results come from production incident records from Microsoft Azure, so the abstract alone does not show how the system performs in other environments.
Key points
- TSGuard is a user-centric multi-agent system for diagnosing incidents in cloud AI workloads.
- It builds domain-specific knowledge bases from historical on-call experiences.
- It uses structured reasoning and iterative trial-and-error to imitate human expert diagnosis.
- In Microsoft Azure incident records, it improved diagnostic accuracy by 19.8%.
- It reduced average verification time by 63.4% versus a sequential execution baseline.
Disclosure
- Research title:
- TSGuard improves cloud AI incident diagnosis accuracy
- Authors:
- Yitao Yang, Yu Deng, Yifan Xiong, Baochun Li, Hong Xu, Peng Cheng
- Institutions:
- Chinese University of Hong Kong, Chinese University of Hong Kong, Chinese University of Hong Kong, Microsoft (Canada), Microsoft (United States), University of Toronto
- Publication date:
- 2026-06-30
- DOI:
- 10.1145/3797149
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with scholarly research without reading academic papers — one filtered digest, every Friday.