Identify top-level agent spans

[javiermtorres]

Description

I would like to know if there is some way of identifying the top-level span used by an agent. With workflows and teams, it is not easy to know how many agents are intervening and when a specific agent's actions are starting.

[Jgilhuly]

Hey @javiermtorres - great question. Depending on your instrumentation, the agents could be shown hierarchically or sequentially in the trace, but either way you can use the same approach. You iterate through spans and follow the parent_id column to move up the trace, checking for openinference_span_kind=="AGENT" to find the root agent span.

We're doing some work on trace level evals now that requires this. I can create an example. What framework are you using for development, if any?

[javiermtorres]

The issue with that approach is that chains or groups use several agents, so the top level span for some agent may not be the topmost AGENT type span, since it could be started in mid-chain. So how could I identify the start of a series of spans issued by the same agent, when taking a complete trace into account?

[axiomofjoy]

@javiermtorres Do you have any particular framework or library you're working with?

It might make sense for us to add metadata for "agent_id": <id> to enable filtering.

[javiermtorres]

True, I hadn't thought of chrono/hierarchically organizing the spans in the trace and find some distinct id or count the ids (if the underlying framework supports unique ids). Using such an attribute would definitely be a plus, since names are probably not unique in most frameworks. We're working with quite a few frameworks, so having a common key for this attribute is really helpful.