name: "wide_research"
description: "Use when the user needs broad parallel research across many independent inputs with a shared output schema."
metadata: { "includeInPrompt": false }
Wide Research / 广域研究
Purpose / 目的
Coordinate one manager subagent that fans out the same research task across many independent inputs and returns a single normalized result set.
协调一个管理者(manager)子智能体,把同一研究任务扇出到多个相互独立的输入上,并返回一个统一的规范化结果集。
When to Use / 何时使用
- The task can be split into many independent subtasks (for example, one company/person/topic per input).
任务可以拆分为多个相互独立的子任务(例如每个输入对应一家公司/一个人物/一个主题)。 - Each subtask should return the same structured fields.
每个子任务应返回相同的结构化字段。 - The user asks for "wide research", "parallel research", or high-volume lookup/screening.
用户要求"wide research"(广域研究)、"parallel research"(并行研究)或大批量查询/筛选。
Do not use this workflow for single-item tasks or when subtasks depend on each other.
单项任务或子任务之间存在依赖时,不要使用此工作流。
Tooling / 工具
Use exactly one manager subagent for the overall operation.
整个操作只使用一个管理者子智能体。
Manager input contract:
管理者输入契约:
operation_brief: one concise sentence.operation_brief:一句简洁的说明。inputs: one independent item per element.inputs:每个元素对应一个独立条目。output_schema: required fields and allowed types.output_schema:必需字段及允许的类型。worker_prompt_template: per-input instructions.worker_prompt_template:针对每个输入的指令。completion_format: exact JSON object shape for the manager's final response.completion_format:管理者最终响应的精确 JSON 对象结构。
Manager output contract:
管理者输出契约:
totaltotal(总数)success_countsuccess_count(成功数)failure_countfailure_count(失败数)resultsresults(结果列表)failuresfailures(失败列表)notesnotes(备注)
Operating Rules / 操作规则
- Deduplicate and normalize the input list before spawning the manager.
在启动管理者之前,先对输入列表去重并规范化。 - Keep the output schema minimal and explicit; avoid optional or free-form fields unless the user asked for them.
输出 schema 保持最小且明确;除非用户要求,避免可选字段或自由格式字段。 - Use one manager coordinator for one user goal; do not fan out multiple sibling root-level subagents.
一个用户目标只使用一个管理者协调器;不要扇出多个平行的根级子智能体。 - Instruct the manager to assign one worker per input and keep each worker scoped to its own item.
指示管理者为每个输入分配一个 worker,并让每个 worker 只处理自己的条目。 - If completeness matters and some inputs fail, retry only the failed inputs, once, when feasible.
如果完整性很重要且部分输入失败,在可行时只对失败的输入重试一次。 - Reply to the user immediately that wide research has started; do not block on completion.
广域研究开始后立即回复用户;不要阻塞等待其完成。 - In the final user-facing output, always report coverage (
success_count/total) and unresolved gaps.
在最终面向用户的输出中,始终报告覆盖率(success_count/total)和未解决的缺口。
【评论】"一个目标只用一个管理者、不扇出多个根级子智能体"是对并行规模的约束,用于控制成本与编排复杂度;"开始即回复、不阻塞等待完成"则是降低感知延迟的交互设计。