Why we label every span instead of a sample
Alex Li
|
Sampling the judge is a rational answer to per-call prices. It leaves most runs without a label, and those are the runs you need to group a pattern or find its cause.
Why we group agent runs by the shape of their execution path
Alex Li
|
The same cause gets investigated by different people on different days. Group every run by its execution path, which tools in what order, and each repeat joins one shape.
Every trace looks green and the agent is still failing
Alex Li
|
Every span returns OK and the user still got the wrong answer. The failures that cost trust don't throw. Give every span a label so they have a name you can count.
Nobody should have to open hundreds of agent traces one at a time
Alex Li
|
At a hundred runs, opening every trace is diligence. Somewhere before a thousand it turns into a sample nobody chose. One picture of all agent runs fixes that.