LLMs, GenAI, RAG, agents, evaluation · insight-0633
Make tools ask for missing inputs instead of acting on values the model filled in
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Applied AI, ML engineer
- Industries
- aviation, e-commerce
Summary: situation, action, reason, boundary
Situation
A user says "rebook me on the next flight" without a date or booking reference, and the rebooking tool requires both.
Action
For side-effecting tools, require that critical values come from the user's words, the session or a lookup the agent performed; otherwise return a needs_input result naming the missing fields so the agent asks a question instead of acting.
Reason
Anthropic's tool-use docs note that when required parameters are missing, some models infer reasonable-looking values rather than ask. A required field in the schema forces a value to exist, not to be known.
Boundary
Booking, payment, messaging and account-change tools.
Symptom or risk
The model fills the required fields with plausible defaults (today's date, the most recent booking), the call is schema-valid, and the wrong itinerary changes.
Exceptions
Low-stakes read-only lookups, where a guessed default costs one extra call and can be shown to the user.
Trade-offs
More clarification turns lengthen conversations; ask for all missing fields at once.
Pitfalls
Making critical fields optional to avoid validation errors, which turns guesses into silent defaults.
Lessons and concepts
Claims and their sources
Anthropic's tool-use documentation says that when a prompt lacks required parameters some models may infer reasonable-looking values instead of asking, and client tools are executed by the application.
Tool inputs are described with JSON Schema, whose required keyword demands that a property be present but cannot say where its value came from.