What did you rewrite after watching an agent use your tool?

I ship a small reminder product with an MCP server, so AI agents can create reminders on a user's behalf.

Two things I learned this week that I did not expect, and I would like to hear yours.

1. The description is the behaviour. My create tool said to lead with a phone call because it is the channel that gets attention. So every reminder the test agent created started with a call, even "email me the invoice on Friday". Calls are also the most expensive channel. I rewrote the description to say the chain is the user's choice, one step or several, any channel, any order, and not to add steps the user did not ask for. The next request behaved. No code changed, only words.

2. A destructive tool needs a safer sibling, and the description has to name it. My cancel deletes the reminder and its history. Pause keeps everything. Once the cancel description said in capitals that it is irreversible and that the agent should pause unless the user clearly said delete, the agent started pausing on anything ambiguous and asking first.

Both felt like prompt engineering, but they are product decisions about cost and reversibility, made in a text field most users never see.

If you have shipped tools for agents, what did you have to rewrite after watching them run? And do you test tool descriptions the way you test code?