What observra telemetry looks like

Synthetic telemetry from three agent sessions at a fictional company, "Atlas" — a clean run, a flagged run, and a failed run, each on a different framework. The session cards below tell each story; search or filter the table, and click any row for the raw event JSON.

Generated sample data (see demo/generate_data.py), not a live agent run — shaped to match observra's real event schema. Framework, model vendor, and agent identity are independent fields, and the sample exercises that: one agent runs Claude on Google's ADK, another splits models by role within a single session.