智能体 / NAVIN RESEARCH
Design Better Tools for AI Agents
Make agent tools easier to select, call, validate, and recover from with narrow schemas and actionable errors.
Make agent tools easier to select, call, validate, and recover from with narrow schemas and actionable errors.
简要回答: Design tools around user intentions rather than low-level API endpoints. A good tool has a distinct name, a concise description, a strict input schema, a small response, and errors that tell the agent what can be corrected.
实施步骤
- Start from real tasks and list the minimum operations needed to complete them.
- Use enums, required fields, examples, and server-side validation.
- Return stable identifiers and compact summaries; store large payloads outside model context.
- Separate read operations from writes and mark consequential actions clearly.
验证清单
- Measure tool-selection accuracy with confusing neighboring tools.
- Fuzz invalid arguments and ensure no side effect occurs.
- Confirm retries are idempotent or carry an idempotency key.
常见错误
- Exposing every backend endpoint as a tool.
- Descriptions that differ only by one vague verb.
- Returning unbounded logs or documents into context.
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。
