Following the Trail of a Rogue AI Agent with OpenTelemetry
When an AI agent does something wrong in production, you can’t reproduce it, yet someone will ask you to explain exactly what happened.
The good news is that collecting telemetry has gotten easier. Spring AI, Strands Agents, Pydantic AI, and other frameworks now emit OpenTelemetry’s GenAI semantic conventions. The catch is that those conventions are still in development and framework defaults differ, so having the data is not the same as being able to trust it. To test that, we’ll play detective in a live demo with Java and Python agents. An agent has deleted rows from a database, and we’ll investigate using traces alone. Whether you build agents or have to answer for them, you’ll leave knowing how far your traces can take you.


