How to read agent outputs
What the structured result of a run contains, why schema validation is a hard gate, and where the full trace lives.
Task agents return typed objects validated against the output schema declared in the agent's descriptor — never free-form text a human has to re-interpret.
Anatomy of a result
Each agent declares its own output schema in the Catalog. The ARO reconciliation investigator, for example, returns a BreakAnalysis: a probable root cause, a confidence score, the evidence the agent gathered with links back to source data, and a recommended action (auto-correct, manual review, or escalate). Your workflow — or you — decides what happens next.
Validation is a hard gate
If an agent's output violates its declared schema, the run fails explicitly. A malformed answer is never silently surfaced as a success, so downstream workflows can trust what they consume.
Every run is traced
Each run records the model used, the ordered tool calls with their inputs and outputs, token usage, cost, and duration. Platform operators can open the full trace in Langfuse, and auditors can later reconstruct the run from the audit log — including the exact Catalog version that executed.
Still stuck?
Tell us what you were trying to do and where it failed — screenshots and run IDs help. A human follows up.