Ever wondered what happens within an Agentic AI Workflow? My colleagues added some new demos to our Playground such as an FAQ Agent for Travel Advisory and instrumented it with OpenTelemetry and #OpenLLMetry from Traceloop. Below is how such a multi-step agentic workflow looks like. We see 👉Every involved agent 👉Prompts to Models 👉Calls to Tasks 👉Decisions 👉Timings and Exceptions
Daniel Braaf
5mo
This will become very important for Agentic Ai! End to end visibility into the entire process from initiation to what changes were made is essential, especially at scale. As i always say before you automate something, make sure to consider what to do if something goes wrong, you can't do that without end to end visibility into the entire process!
I love that topic and believe that OTel instrumentation is really important for production-ready AI agents. A few additional tips based on my practical experience: - You likely don’t want to capture LLM input/output or the thinking output on prod -> make it configurable - Build a mental model when instrumenting your agent. What do you care about? -> Manually instrument those parts to avoid Trace explosions - Don’t rely on span attributes if you need accurate usage tracking/billing -> Traces/Spans might be sampled out or lost