Tracing and evaluating an agent before every prompt change

An agent that drafts letters to real clients needs more than a demo that looks right. This work adds two things to InvoiceChaser.

The plan

• Per-step tracing — every prompt, tool call, tool result and human decision recorded for each run, so a bad draft can be debugged from the trace instead of guessed at.
• A golden-set eval — a fixed set of disputes at each escalation level, with checks on tone, which tools were called, and that no draft skipped approval. It runs on every prompt or model change.

[Update this page with the tracing tool, the eval results and a screenshot once it ships.]