A single agent trace is useful when one run fails. It is not enough when the same failure pattern appears across many users, tools, prompts, and retries. The Comet article on Opik agent diagnostics points at the next layer: agent observability has to move from reading one transcript to debugging a population of traces.
The shift matters because production agents fail in distributed ways. A bad tool argument, a stale prompt assumption, a cost spike, and a loop that retries the same weak plan can all look like ordinary message history unless the traces are grouped by behavior. Diagnostics are the layer that turns raw trace logs into a triage surface.
For New Runtime, the useful pattern is not “store every trace forever.” It is: capture enough structured trace evidence, cluster recurring failure modes, connect those clusters back to prompts/tools/runtime boundaries, and keep the repair queue auditable. The trace is source material; the product is the decision about what to fix.
The practical consequence is that agent platforms need observability objects that are closer to issue trackers than dashboards: failure classes, affected workflows, prompt or tool versions, reproduction links, and repair status. That is where debugging becomes operational instead of artisanal.
New Runtime synthesis: once agents are normal software actors, trace diagnostics become part of the release process. They tell the operator where the next prompt, tool schema, evaluator, or runtime gate should change.