Module 2 · Operating agents · Lesson 5
Observability and cost
Traces that show what an agent did and why it was slow or expensive, OpenTelemetry and eve Agent Runs in practice, and token, latency and cost budgets per case, with prompt caching and model routing worked through in numbers.
By the end of this lesson you can
- Explain traces, spans and attributes, and map an agent run onto them using the OpenTelemetry GenAI conventions
- Enable tracing in eve and the Claude Agent SDK, and control which content reaches each destination
- Compute the cost of an agent run with and without prompt caching, and predict when caching stops helping
- Set and enforce token, cost and latency budgets per session and per case
- Use traces to find the step that dominates cost or latency, and choose between caching, routing and redesign