Observing an agent

3 posts, in reading order. Drafts are previews until they ship.

1 Observability is built in; where it goes is a layer draft Spans, metrics, and logs are Effect built-ins, not a vendor SDK you thread through your code — and one layer at the edge decides the backend, so telemetry-off costs nothing.
2 Name spans for humans, query them by attribute draft What to trace in an agent loop — and why the stable attribute you put on a span outlives the name you give it, for dashboards and for evals alike.
3 Your eval has no database; the trace is the dataset draft Once the agent loop is fully traced, evals need no separate store — the runner discards its own return value and rebuilds the scorecard by reading the span tree. Tokens, steps, and cost come from the same trace the dashboards do.