Traces and Observability

In one line

A record of everything an AI system did on each run: what it read, what it decided, what it called, what it cost.

Picture it

A flight recorder. Nobody reads it on a good day; on a bad day it is the only thing that explains what happened.

Under the hood

One trace per run capturing the plan, each retrieval and what it ignored, every tool call with inputs and outputs, rejected alternatives, and the cost of each step. Logs tell you what the agent did; traces tell you why. This is what answers 'why did it do that' six months later.