Teams evaluate agents by trying a few prompts and nodding. Then they change the model, the prompt or a tool, try a few more prompts, and nod again. Nobody can say whether the agent got better. The fix is an evaluation suite, and it is cheaper than it sounds.
Build the suite from real cases
- Collect 50 to 200 real inputs the agent will face. Not synthetic. Real orders, real tickets, real listings.
- For each, record what a good decision looks like. Often this is what your best human did.
- Write a scorer. Exact match where possible, a rubric where not.
- Run every change through it. Prompt edit, model swap, new tool, all of it.
What to measure
- Decision accuracy: did it choose what the human would have?
- Guardrail hits: how often did a check fire? Rising hits mean the agent is drifting.
- Cost per decision: tokens and tool calls.
- Latency: customers notice.
The eval suite is the spec
Once it exists, the eval suite becomes the contract between the model team and the business. The business says "it must get these cases right". Engineering says "it does, here is the score". Arguments about whether the agent is "good enough" turn into a number.
Do this today: pull 50 real cases from last month, label them, and run your current agent. The score will surprise you.