Applied AI Systems

A tool definition is a contract, and the model reads it literally.

Design the tools, schemas, and return shapes that decide what a system can actually do.

The boundary between a model that answers and a system that acts is a set of tool definitions. Each one declares what can be called, what it needs, what it returns, and what happens when it fails. That surface determines behavior more directly than any instruction wrapped around it.

This is where capable systems behave badly. A vague parameter description produces confidently wrong arguments. An oversized tool set makes selection unreliable. An error returns as prose the model reinterprets rather than as a state it can act on. The instinct is to rewrite the prompt; the fault is usually in the contract.

Tool & Function Calling treats that surface as engineering. It gives the team well-scoped tools, schemas precise enough to constrain arguments, return shapes that carry state rather than narration, and failure paths a system can respond to. Prompts are one mechanism here, not the headline.

Let’s get going

  • Start with the call that goes wrong most — Pick the tool most often invoked with bad arguments and tighten its schema and description before touching anything upstream of it.
  • Scope tools to decisions, not endpoints — A tool should map to something the system decides to do. Exposing an API surface one function at a time makes selection harder for no gain.
  • Return state, not prose — Structured results and explicit error states let the system respond. Narrated failure gets reinterpreted, and reinterpretation is where side effects come from.

Outcomes

  • Reliable invocation — Tools are called with valid arguments more often, and selection stays workable as the tool set grows.
  • Failures that stay contained — Errors surface as states the system handles rather than as text it interprets and acts on.
  • A capability surface that can change — Adding, removing, or revising a tool is a contract change with a known blast radius rather than an unpredictable behavior shift.