How AI helped us ship updates in a Linux distro
summary
Ymir is an end-to-end agentic service that produces merge requests to backport CVE fixes into RPMs but never auto-merges them; agents run in OpenShift pods using the BAI framework and interact with GitLab/Jira through an MCP gateway so no credentials leak. The system uses 70 tools and about 1,700 lines of prompts, runs dry‑runs and frozen end‑to‑end tests, stores traces in Phoenix and a custom UI, and sometimes takes hours but produces reviewer-ready MRs.
AI-written summary · about
chapters (9)
- 0:00AI-driven Linux updates
- 1:02Rising CVEs and AI
- 2:25Ymir automation system
- 7:01Ymir architecture
- 10:01AI system design challenges
- 18:55Observability in AI automation
- 21:25Testing AI-driven systems
- 24:16Real-world impact and future
- 27:08Q&A
related talks
31mAgents at Work: How GitLab Is Redefining AI Adoption for Engineering Teams
18mUsing AI Across the Dev Lifecycle
27mIntegrating Agentic AI Into Internal Developer Platforms To Minimize Errors and Downtime
17mFighting AI with AI
21mAI-driven optimization of CI/CD pipelines
28mAI for Everything After Code
42mFrom Logs to Fixes: Teaching AI to Maintain Software
23mGitOps Meets AIOps: Building Reliable AI Pipelines With Argo