An agent that works twice is not necessarily an agent that works. Before you let one near real work, run it through a few checks. It takes about half an hour and saves you the awkward kind of mistake.
1. Try it on something you already know the answer to
Pick a past task where you remember the right result. Run the agent on it and compare. If it gets a task you already solved wrong, you have learned something cheaply.
2. Give it a messy input
Demos use clean examples. Real work does not. Try a half-finished note, a missing detail, an odd format. You are looking for what it does when information is missing. Good setups ask or flag the gap. Weak ones quietly fill it in.
3. Ask it to show its reasoning
For anything that matters, have the agent say which parts of your input it used and what it assumed. Wrong assumptions are much easier to spot when they are written out.
4. Run it more than once
Use the same input three times. If the results differ a lot, the instructions leave too much open. Tighten them until the output is steady.
5. Decide what a human still checks
Even a good agent gets a final look from you. Decide in advance what that look covers: the numbers, the names, the tone. Write it at the end of the instructions so you do not forget.
Keep a short list of failures
Every time the agent gets something wrong, add a line to the instructions that prevents it. After a few rounds the skill gets sturdier, and you have a written record of what to watch.
The free starter kit includes a one-page version of this checklist.