Observability model
Two things carry the record of what happened: the execution (the run itself, and what each step produced) and the audit log (who caused it). Neither is a full distributed-tracing product yet. Here's what's real.
Execution records
Every run is backed by an execution record. Each step in the execution record carries its status, inputs and outputs, any human approvals, agent runs, the resolved surface (either output or wait), and loops summaries for loop steps. This is the primary way to inspect a run in progress in the Console.
A step waiting on approval carries an outcome once it resolves (approved, rejected, or timed_out), alongside pending while it's still open.
Execution, model-call, and audit records
Administrative actions (such as deploys, resource renames, OAuth credential callbacks, and webhook secret regeneration) land in the append-only audit log, attributed to the user who triggered them, while runtime activity is recorded across workflow executions, agent execution events, and model-call accounting records. See Governance and Identity and audit attribution for what's attached to each record.
What we haven't formalized
While the Console provides a workspace Health view with run, token, and cost metrics, a public external trace-event schema and pluggable trace export remain on the roadmap. A fixed catalog of trace event names (something like run.start / step.ok / tool.call) and per-call cost/token attribution aren't a committed, documented schema yet. If you're building automation on top of execution data today, treat field shapes as something to confirm with us rather than something to assume is stable. The Console UI in particular changes without notice. Pluggable trace export (OpenTelemetry, Datadog, Honeycomb) is on the roadmap, and a stable, versioned event schema is a prerequisite for it.