Traces and audit in your own stack.
See every turn an agent runs, in the same shape across every harness: a built-in trace console for day one, OpenTelemetry export for the stack your team already watches, and an immutable audit log for your auditors.
Console, export, and the immutable trail.
Every org gets a session-by-session trace view at four fidelity levels, with tokens, cost, duration, and models. No setup; it is simply on.
Standard OpenTelemetry GenAI traces - invoke_agent, chat, execute_tool spans with token and identity attributes - delivered to Langfuse, Datadog, Grafana, or any OTLP/HTTP collector you run. Open standard; no lock-in.
Who did what, when - enforced immutable at the database, exportable over API and CSV behind a dedicated scope. Exporting the log is itself an audited event.
agent runtime ─▶ traces · audit events ─▶ { console · your OTLP backend · your export tooling }
Designed for the security review.
Metadata by default
Exports carry operations, models, token counts, tool names, and identity attributes - never message content unless you switch your own endpoint into content mode, and content only ever flows to your endpoint.
Deep links from your product
Each turn surfaces its trace id on the session stream, so your product can link a conversation straight to its trace in your backend.
Uniform across harnesses
Traces are synthesized by the platform, not by the harness - the same span shape whether the agent runs on Claude Code, Goose, or Codex CLI.
Endpoints held to your bar
HTTPS only, credentials in encrypted headers, private addresses rejected, session-coherent sampling. Export enablement is configured with you - talk to us.