The talk works through one agentic app end to end, in the order you would actually hit these problems:
- A running example — a real agentic app, not a toy prompt
- Building a cost dashboard with Claude Code, so the spend is visible at all
- Caching tactics, including the ones that quietly make it worse
- A run of experiments to push the number down
- An eval harness, so a cost cut gets checked against quality instead of assumed
- What to take back to your own agents