Agent Observability access without access to raw agent content

Raw agent content — prompts, completions, tool inputs and outputs, transcripts, and judge rationale — now has its own permission. A user without it still gets the full operational product: summary metrics, span structure, monitors, and alerts.

Until now, "who can use Agent Observability" and "who can read everything your users typed" were the same question, so the audience was capped at whoever cleared the second bar. A platform team can now own agent reliability and setup for the whole organization without ever opening a conversation.

Configure it in the custom role policy editor here. Details in custom role recipes.