ML//agent//tool use

Tool use is a model's ability to request operations such as search, code execution, file access, database queries, and browser interaction, which a surrounding program then carries out, so that the model can act beyond producing text. Function calling is a common structured transport for those requests.

Tool use is a model's ability to request operations such as search, code execution, file access, database queries, and browser interaction, which a surrounding program then carries out, so that the model can act beyond producing text. Function calling is a common structured transport for those requests.

The model does not execute the implementation. It emits a call that an agent harness may validate, reject, transform, or execute.

Tools extend more than capability: read tools determine what the model can learn about the environment, while write tools determine what it can change.

A tool can do exactly what the model does badly. A calculator or a code interpreter gives exact arithmetic where the model would guess digits; this delegates a step to a deterministic engine and makes the numbers right, without making the system reason symbolically about anything else (neurosymbolic AI is the fuller pairing, where a checker judges what the model proposes).

Reliability depends on choosing the right tool, supplying valid arguments, interpreting observations, and recovering when reality disagrees with the plan.

Giving a model a tool is not enough. The tool interface must make the correct use legible and the runtime must make incorrect use bounded, starting with giving each tool only the authority the operation it represents needs (minimum privilege).