Safety Evaluation / NAVIN RESEARCH
Build Tool-Call Evaluations for AI Agents
Test tool selection, arguments, sequencing, authorization, recovery, and unnecessary calls with executable fixtures.
Test tool selection, arguments, sequencing, authorization, recovery, and unnecessary calls with executable fixtures.
Short answer: A tool-call eval should assert the intended operation and state transition, not demand one exact natural-language trace. Use fake or sandbox tools, seed controlled state, and check required calls, forbidden calls, argument properties, and final state.
Implementation steps
- Create fixtures for common and boundary tasks.
- Record allowed, required, and forbidden tool effects.
- Inject typed errors and stale-state responses.
- Score both completion and efficiency.
Validation checklist
- Test write idempotency and duplicate calls.
- Verify denied operations stay denied after retries.
- Review near-miss arguments that pass schema but violate business rules.
Common mistakes
- Exact-match grading of call order when order is irrelevant.
- Testing against production systems.
- Ignoring unnecessary expensive calls.
Practical decision
Ship the smallest design that passes the checks above. Record the model, prompt, tool, dataset, and runtime versions used in testing so later changes can be compared rather than guessed. Recheck the linked documentation before relying on provider-specific limits, model names, prices, or preview features; those details change more quickly than the engineering principles in this note.
