VCN #38: Glass Box — Trace Agents with OpenTelemetry
Build the observability layer for agents from OTel primitives. No SaaS.
Agents fail quietly. Tool calls silently skip. The model returns a hallucination. The trace would have told you, but you didn't ship one, because LangSmith costs money and Braintrust costs money and Helicone costs money and you decided to "add observability later."
Later is now. And later is free.
VCN #38: Glass Box is a builder night where we wire OpenTelemetry into an agent stack from primitives. No managed dashboard until you can read a trace by hand. OpenInference is the de-facto plug-in layer (the OTel GenAI semantic conventions are still in development); we use it because Arize ships instrumentations for OpenAI, Claude Agent SDK, LangChain, DSPy, CrewAI, Bedrock, Anthropic, Google GenAI, LlamaIndex, and a dozen others. Anthropic instrumentation shipped v1.0.3 in May.
Format:
Wrap one LLM call. Three lines of OpenInference. Spans appear.
Boot Langfuse locally.
docker compose up. Self-hostable OSS. Trace UI live in 90 seconds.Read the trace by hand. Span tree, attributes, events, exceptions. Find the failing node.
The managed contrast. Same OTel spans piped into Braintrust. Same data, two views. Which one do you actually need?
By 10pm your agent stack has spans, traces, a working dashboard, and a baseline score. Tomorrow's regression has somewhere to land.
Builders only. Bring an agent that already works (or one that doesn't — both useful).
Doors 7pm. Talks 7:30. Frontier Tower Floor 9.
Hosted by Vibe Coding Nights: Rayyan Zahid (Immersive Commons), Michalis Vasileiadis (Otto / GSD2.0), Eric Mockler (logistics, F11 Health & Longevity), Devinder Sodhi (Tower Lead, Frontier Tower).
RSVP if your agent currently fails in the dark.