April 29, 2026
Standard? More like stan-dardly chaos
Lessons from Building an OTel Normalizer for GenAI
Everyone said the tracking standard was settled — then builders found total chaos
TLDR: groundcover says building one clear view of AI app activity was far harder than promised because every tool reports things differently. Commenters were split between smug laughter and real frustration, with many saying this proves tech "standards" often sound better in marketing than in real life.
The big mood around this post? "So the standard isn’t really a standard". Engineers at groundcover thought they were plugging into one neat shared system for tracking what artificial intelligence apps are doing. Instead, they say they walked into a messy group project where everyone labeled the folders differently and swore they were following the same rules. The community reaction was a mix of vindication, eye-rolling, and dark comedy: plenty of developers basically replied, "Welcome to software, where standards are just strongly worded suggestions".
The hottest takes came from people arguing over who’s really to blame. One camp said this is exactly why companies need tools that smooth over the chaos behind the scenes. Another camp was much less forgiving, saying the whole industry loves to brag about openness until it’s time to make things actually match. A few commenters seemed almost delighted that the shiny promise of easy plug-and-play observability got exposed as a patchwork of mismatched names, weird nesting, and provider quirks. In plain English: different app-building tools, different AI companies, and different tracking kits all describe the same event in different ways, which makes comparing anything a headache.
And yes, the jokes wrote themselves. People compared it to everyone bringing a different charger, or showing up to a costume party dressed as the same character from totally different movies. The funniest running gag was that the real universal standard in tech is confusion, with one recurring sentiment: if you need a "normalizer" this badly, the "normal" part may have already lost the plot.
Key Points
- •groundcover built a GenAI telemetry normalizer to create a single canonical view across multiple OpenTelemetry-compatible SDKs and LLM providers.
- •The article says real-world GenAI telemetry is inconsistent despite OpenTelemetry semantic conventions, with differences in attribute names, structures, and provider-specific behavior.
- •groundcover supports two telemetry collection paths for AI observability: SDK instrumentation via OTLP and eBPF-based interception of LLM HTTP calls.
- •The article identifies three factors that shape telemetry variability: instrumentation SDKs, orchestration frameworks, and LLM providers.
- •Frameworks such as LangGraph, CrewAI, and Pydantic AI can change span hierarchy, metadata, and message serialization even when the same provider and SDK are used.