Agentic Backstage: How To Manage an AI Software Catalog
summary
Roadie details building a catalogue-driven agent platform: open sourced vector database and RAG layer, produced MCP servers and ingestion providers, and used Claude to generate providers in a day (one experiment in 38 minutes). Agents now outnumber users about 100:1, and the system resolves roughly 80% of on‑call alerts without human intervention. The catalog is refreshed from GitHub/AWS/Datadog/PagerDuty/Argo, relationships are inferred algorithmically or with lightweight models, and actions run through the scaffolder.
AI-written summary · about
chapters (11)
- 0:00Agentic Backstage intro
- 2:04Two years of AI experiments
- 3:49What agents need
- 7:23Building software graph
- 10:30Catalog relationships
- 13:20What didn’t work
- 14:56AWS permission mapping
- 15:49Delivering usable data
- 17:34Agent action enablement
- 19:40Live incident demo
- 20:46Q&A
related talks
30mBuilding Robust, Secure LLM and Agentic AI Workflows
29mBuild an enterprise Agentic AI platform at Powercor
17mTransforming Robotics Service with Agentic AI on
22mAgentic AI for Engineering Leaders | Strategy & Execution Guide
33mAsk Me Anything: AI Agents, Skills, Career & Growth
06mKeynote: Building Agents at Scale
15mThe Responsible Path to Hyperproductivity with AI Agents
40mAI Agents in Practice