From Spans to Agent Scenario: Why We Added a GenAI View of the Trace

The Proto Observability Platform added a GenAI view for traces, addressing the difficulty of reconstructing agent logic from a simple tree of spans. The view has four modes: a feed showing operations in chronological order with timing, a step graph that collapses repeated operations to reveal the agent's loop, a call graph that shows each individual invocation with its unique arguments and status, and a chat mode that reconstructs the user's conversation from potentially repeated message history. The developers encountered several real-world issues: full conversation histories were sent with each model call, leading to duplicates; transport spans could distort timing if summed naively; structured tool results might contain errors even when the operation status appeared successful, so the error model was separated from the display model; and token consumption could be cumulative or per-operation, requiring careful semantics detection before aggregation. The representation intentionally hides details not relevant to the question each mode answers, while raw attributes remain available for verification.