Analysis agents

Design

Bounded, read-only agents that reason over the spans that actually ran.

#Evidence-linked, not inferred

Analysis agents read the trace; they do not sample it or infer beyond what it can support. Each produces a bounded recommendation carrying the evidence that produced it, so a reviewer can agree or refuse on the facts. Agents are workers, not request-time calls.

Cost anomaly detection — spend drift caught at the span.
Prompt and caching optimisation — verbose or near-duplicate prompts.
Quality and drift monitoring — distribution change over time.