记忆与上下文

How Tool Search loads definitions on demand, affects prompt caching, and differs from MCP connections and ARD resource discovery.

解释 Tool Search 如何按需加载工具定义，为什么它会影响 Prompt Cache，以及 MCP、Tool Search 和 ARD 分别解决什么问题。

Use When Agents Slow Down, Reflexion, Hyperband, and agent evaluation to compare continued revision with independent attempts under reproducible budgets.

结合 When Agents Slow Down、Reflexion、Hyperband 和 Agent 评测，说明怎样比较连续修改与独立尝试，并设计可复现的预算分配实验。

结合 OpenAI Agents API 和 Anthropic Managed Agents，解释 Session、Context Window、Compaction 与持久状态怎样支撑跨上下文窗口的长任务 Agent。

How OpenAI Agents API and Anthropic Managed Agents use sessions, context windows, compaction, and persistent state for long-running agents.

RTK filters test, Git, search, and build noise before command output enters model context, applying context engineering at the tool-output boundary.

介绍 RTK 如何在命令输出进入模型上下文前过滤测试、Git、搜索和构建噪声，把 Context Engineering 做到工具输出边界。

VEP/1 turns images into short evidence packets. Provider fallback, caching, and credential storage manage cost and reliability for text agents.

通过 VEP/1 把图片转换成短证据包，并用 Provider fallback、缓存和凭据存储控制文本 Agent 的视觉输入成本与可靠性。

Break down planning, execution, evaluation, state management, and stopping rules to keep multi-step agent tasks controlled and traceable.

把 Harness 作为一种工程模式，拆解计划、执行、评估、状态管理和终止判断，说明 Agent 多轮任务怎样保持可控和可追溯。