Every vendor now demos an agent that finds the root cause and opens the PR. Almost nobody talks about what that agent is reading — and that's the part that decides whether self-healing is real or a very expensive way to break production faster. Humans tolerate bad telemetry. We squint at an inconsistently named metric, we know which dashboard lies, we carry the tribal context that isn't in any span. An agent has none of that. It acts on exactly what the pipeline says, which makes the data contract — not the model — the binding constraint on autonomy. This talk makes the case that OpenTelemetry is that contract, and gets concrete about it. The GenAI semantic conventions (invoke_agent, execute_tool, gen_ai.*) now let you observe agents as first-class citizens rather than opaque HTTP calls. The same conventions, pointed the other way, are what let an agent reason across your estate without a bespoke integration per tool. Then the honest half: where this breaks today, what the conventions still don't cover, and a staged trust model for moving from suggest → propose → act, with a blast radius you can defend in a post-mortem.
Judith Redi is VP of Engineering at Dash0, an OpenTelemetry-native observability company. She leads the engineering organisation building the platform's ingest, storage, and analysis layers — which means she spends a lot of time looking at what real OTel pipelines do under load, and at what breaks when teams migrate onto them. Before moving into industry engineering leadership, Judith spent a decade in research: a PhD from the University of Genoa on machine learning for image quality assessment, postdoctoral work at Eurecom, and years as an Assistant Professor in TU Delft's Multimedia Computing group, with a Veni award and an IBM Faculty Fellowship along the way. She is a long-standing advocate for gender diversity in engineering.
