Demand-Driven Context: What 300 Incidents Reveal About Tribal Knowledge
summary
An AI agent is run against ~300 real e‑commerce incidents and repeatedly fails until failure is used as a signal to document only what’s missing, a method called Demand-Driven Context. The pipeline uses RAG with Opus 4.5/5, clusters incidents into ~20 types, and runs nine DDC cycles until convergence at 46 entities. A demand_context skill makes the agent list missing vocab, systems, or region rules; a propose_knowledge flow opens PRs to a reviewed KB so context evolves without rotting.
AI-written summary · about
chapters (12)
- 0:00Introduction and context
- 1:47The problem with incident response
- 4:47First failed attempts at AI agents
- 7:50The difficulty of documentation
- 10:33Learning from AI failures
- 13:17Introducing demand-driven context
- 15:48Analyzing the demand logs
- 18:34Validating the framework
- 26:04Preventing knowledge rot
- 27:44Continuous knowledge improvement
- 34:38Summary and key takeaways
- 36:34Closing thoughts and resources
