安全评估 / 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.
简要回答: 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.
实施步骤
- 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.
验证清单
- Test write idempotency and duplicate calls.
- Verify denied operations stay denied after retries.
- Review near-miss arguments that pass schema but violate business rules.
常见错误
- Exact-match grading of call order when order is irrelevant.
- Testing against production systems.
- Ignoring unnecessary expensive calls.
实际决策
先交付能够通过上述检查的最小设计。记录测试中使用的模型、提示词、工具、数据集与运行时版本，让后续变化可以被比较而不是被猜测。依赖供应商的限制、模型名称、价格或预览功能之前，请重新检查链接文档；这些细节变化往往快于本文的工程原则。
