tool selection

When an agent has many tools — a search engine, a calculator, a database query, an email sender — it must first decide which one fits the step in front of it. Tool selection is exactly that choice: matching the current need to the right instrument, the way you reach for a screwdriver rather than a hammer once you see the screw. Pick wrong and the step fails or wastes a turn; pick right and the task moves forward.

The model makes this choice from the tool descriptions you provide: it reads each tool's name and summary, compares them against what it is trying to do, and emits a call to the one that matches. Selection gets harder as the toolbox grows — descriptions start to overlap, and a long list crowds the context window. Common fixes are clear, distinct descriptions, grouping or filtering tools by relevance before the model sees them, and keeping the active set small.

Selection and arguments are separate failures, and it helps to keep them apart when debugging: an agent can pick the perfect tool yet fill in the wrong inputs, or choose a clumsy tool that still happens to work. Watching which mistake it makes tells you whether the fix belongs in the tool descriptions or in the examples you give the model.