Once again, this breaks core assumptions in traditional observability. First, traces must be mutable, so that you can save annotations and query them alongside other metadata. Supporting updates on traces at "agent-scale" makes the agent-tracing database problem even more challenging. Second, the users who annotate are rarely developers, and so they benefit from UIs that simplify the data they must annotate into its simplest components.
...
the users who annotate are rarely developers ... maybe they are evals persona -- domain experts... that simplification of the tooling is one reason why Hamel Husain suggests building your own human eval tools. Because you can build for the other personas. UI-UX. Also captured in the ideas that it can be helpful to represent data in a more user-friendly review format