What are you using to trace an agent across queues and tool workers?
We are trying to trace one agent run across an API gateway, queue, vector lookup, model call and tool worker. Each service reports healthy, yet one request can still fail because parent child context disappears at the queue, sampling drops the useful span, or a malformed tool argument triggers a retry. Basic logs have hit their ceiling. Five healthy services can still create one broken request.
We're testing Braintrust for end-to-end span trees, model and tool metadata, trace search and trace level scoring. We also looked at Langfuse since it's openn source and self hostable but context propagation across async boundaries and trace search at our volume were concerns. I need context propagation that survives async handoffs, span links for retries, queue latency and a sampling policy that keeps error paths without exploding cost. How does this compare with established observability stacks in real agent systems?