Recommended Langfuse setup for WDK workflows + DurableAgents?

[simpautus]

I’m integrating Langfuse tracing into WDK workflows and DurableAgents and could use a sanity check on the intended approach.

I’m registering the Langfuse span processor via Next.js instrumentation, and I do see workflow traces arrive in Langfuse. Two issues though:

1. Workflow traces are too noisy / infra-heavy
   Right now I’m getting a large number of WDK infrastructure spans with metadata that’s hard to interpret. What I’d like instead is to define spans/traces explicitly inside the workflow code (and ideally control span “types” / naming / tags), so only the meaningful steps show up with labels I choose.

2. DurableAgent emits no traces
   For DurableAgent I’m currently seeing nothing exported, even though I believe the tracing config is wired correctly. Ideally I’d use the AI SDK integration there to capture the comprehensive model/tool/request metadata without manually adding spans everywhere.

I’ve tried a couple custom approaches, but keep running into issues that seem related to the WDK sandbox constraints.

Questions:

• Is DurableAgent telemetry expected to work today, or is it not implemented yet?
• For workflows, is there a recommended pattern to get more targeted/manual span control (e.g., suppress infra spans, or prefer user-defined spans)?

Any pointers, docs, or examples would be appreciated.

[pranaygp]

Have you seen the observability traces built into workflow by chance? When using DurableAgent, the workflow observability built into should give you all the info you're looking for in the spans. we've done some work towards improving the ovservability but definitely lots more left to do

as for langfuse setup

1. the otel traces/spans emitted from the workflow client are primarily meant to track performance of workflow and the infra, so that makes sense. I guess we could have a flag that lets you disable these otel spans? They're going to be useful when emitting to datadog/etc. but may not so much on langfuse
2. DurableAgent is missing telemetry rn. We'll need to reconcile this with AI sdk (which is also focussed on improving the observability story rn) so that there's consistent telemetry coming from both. This is a notable shortcoming rn

[simpautus]

Thanks, makes sense.

For DurableAgent, I’ll wait for the built-in telemetry support. Once that’s available, I’ll try adding my own clean Langfuse spans on top. I’m still a bit unsure what the best approach looks like with the workflow constraints, so I’ll experiment then.

In the meantime I’ll take a closer look at the built-in WDK observability and whether it can be extended with custom metadata. Ideally I’d like to keep a single place for unified traces/observability, if possible.