Your Agent Did What? Forensic...
summary
Claim: OpenTelemetry becomes the single convergent standard and the genainormalizer processor at the collector layer bridges incompatible frameworks. Demo builds a local stack: a malicious goose deletes free-plan rows, three SRE agents (Beaver: Python + OpenInference, Otter: Python + OpenLLMetry, Capybara: Java + LangChain4j), a Postgres DB, MCP servers, Collector → Jaeger/Prometheus, and Perseus UI. OpenInference spans get translated to OTel GenAI semantic conventions by the GenAI normalizer, tiny tool-call spans reveal the culprit, and missing context propagation hides the full why.
AI-written summary · about
chapters (14)
- 0:00Welcome and introduction
- 1:44LLM application terminology
- 3:32Agent loop and complexity
- 5:21Instrumentation fundamentals
- 7:50OpenTelemetry for LLM apps
- 12:05Demo setup and scenario
- 14:54Malicious goose in action
- 16:20SRE agents investigate
- 17:44Telemetry analysis
- 20:35Forensic analysis findings
- 21:18Context propagation importance
- 22:22Semantic conventions today
- 23:04LLM as judge evaluation
- 23:45Key takeaways summary
