推理 / NAVIN RESEARCH
Reasoning With Tools: Separate Planning From Execution
Let models choose actions while deterministic code validates, executes, and records each operation.
Let models choose actions while deterministic code validates, executes, and records each operation.
简要回答: The model may propose a plan and tool calls, but the application owns permissions, validation, execution, retry policy, and stopping. This separation keeps failures inspectable and prevents natural language from bypassing control logic.
实施步骤
- Expose a narrow set of task-relevant tools.
- Validate each call against schema, user scope, and current state.
- Return concise results with stable identifiers and explicit error types.
- Stop on success, budget exhaustion, policy denial, or repeated non-progress.
验证清单
- Simulate stale state between planning and execution.
- Test partial tool failure and duplicate requests.
- Confirm the agent cannot manufacture a tool result in plain text.
常见错误
- Executing a prose plan as code.
- Returning stack traces or secrets to the model.
- Allowing endless plan-revise loops.
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。
