推理 / NAVIN RESEARCH
A Context Engineering Checklist for AI Agents
Curate instructions, tools, evidence, state, and history as a limited resource on every model turn.
Curate instructions, tools, evidence, state, and history as a limited resource on every model turn.
简要回答: Build context from explicit layers: stable policy, task contract, current state, relevant tools, retrieved evidence, and a compact recent history. Give each layer an owner and a removal rule. More context is useful only when it changes the correct next action.
实施步骤
- Reserve the beginning for stable high-priority instructions.
- Select tools on demand and avoid loading large catalogs blindly.
- Retrieve evidence with provenance and a relevance threshold.
- Compact completed work into decisions, artifacts, and open questions.
验证清单
- Ablate each context layer and measure task impact.
- Detect conflicting instructions before sending the request.
- Track tokens by layer to identify growth.
常见错误
- Mixing untrusted documents with system instructions.
- Keeping stale tool definitions.
- Summarizing away constraints or unresolved risks.
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。
