子 Agent 该用多大力气想？从 Claude Code 2.1.316 的 CLAUDE_CODE_SUBAGENT_EFFORT，聊多 Agent 编码的 effort 分层
短版摘要（供短帖渠道使用） 10 月 31 日发布的 Claude Code 2.1.316 新增了 CLAUDE_CODE_SUBAGENT_EFFORT ：当没有其他设置指定时，子 Agent 用这个档位运行，而不再默认继承主会话的 effort。这看起来是个小开关，但对多 Agent 编码的成本影响很大：主会话开 xhigh 做规划，搜索、读代码、跑测试的子 Agent 往往用 medium 甚至 low 就够了。本文梳理 effort 的完整优先级（环境变量 > 会话内选择 > 启动参数 > 设置文件 > 模型默认），子 Agent frontmatter 怎么写，以及一套“主会话高、子任务低、审查单独拉高”的分层配置。 用 Claude Code 跑稍微大一点的任务，你会发现真正花 token 的往往不是主会话，而是它派出去的一堆子…
Claude Code 2.1.316 introduces CLAUDE_CODE_SUBAGENT_EFFORT, a setting that determines how much computing power sub-agents should use when no other settings are specified. The default is medium, but it's recommended to set it lower (e.g., medium or low) for most tasks to reduce cost. This article explains the configuration hierarchy for effort, including environment variables, session selections, launch parameters, setting files, and model defaults.
It also provides a tiered configuration example, where the main session uses high effort for planning, sub-agents use low to medium effort, and a specialized reviewer agent uses high effort for deep review. The article emphasizes the importance of setting effort at the right level for each task to optimize token usage and cost.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
