Tracing

Every run is a Cencori session, and every session is traced. There's nothing to set up.

Open Observability → Traces in your project:

• Timeline of the run: each model call, tool call and wait, on one time axis. Long idle gaps (a person deciding on an approval) are folded so the work stays readable.
• Tree of agents: handoffs and agents called as tools, with their own sessions linked to the caller's.
• Each turn: the exact messages the model saw, the tools it had, its settings, the output, tokens, cost, time to first token, and which provider served it (and whether a fallback was used).
• Events: the raw session event log.

Filter traces by status (active, waiting, ended, failed) and by source (Agents SDK, Arcie, embedded agents...). The agent's name comes from your Agent, so a team of agents reads by name.

Add your own labels with metadata:

await run(support, input, { metadata: { ticket: 'T-4821', channel: 'email' } });

Each model call in a run is also in Observability → Request logs, with the prompt, response, latency and cost, linked back to its trace.