下面很酷的视频是一个开发者把300个Kimi K2.6智能代理同时运行的状态。

屏幕上每一条跳动的线都是实时传递的信号，像大脑神经元同步激活，双向传输。这不是概念图，是模型处理请求时真实运行的状态。

300个代理并行跑，协调4000个步骤。画面很酷，但有个问题随之而来：谁来指挥这300个代理？

如果让每个代理自己决定做什么，结果就是混乱。如果让最聪明的模型来干所有活，包括指挥和执行，那么结果就是账单爆炸。

有团队之前就是这么干的。他们把全部工作交给Claude Opus 4.8，调研、写稿、分析、批量处理，全用它。每个月光API费就花6.2万美元。Opus本身没问题，它的推理能力、处理模糊问题的能力确实顶尖，但它天生就不是用来做批量执行的。

Anthropic给Opus 4.8做的Dynamic Workflows功能，本质是让它管理复杂任务，编排上百个并行子代理，不是让它自己当子代理干所有活。

他们调整了分工：Opus 4.8当大脑做策划编排，Kimi K2.6 Agent Swarm当手做执行。输出质量没变，规模没变，每月成本直接降到129美元，省下5.42万美元，一年就是65万美元。

分工逻辑很简单：

• Opus 4.8是战略家：做计划、推理、判断、整合复杂信息、把控质量，只处理需要判断力的活
• Kimi K2.6是执行者：同时跑300个专业子代理，协调4000个步骤，批量处理，输出真实文件，SWE-Bench Pro得分58.6%，每百万tokens只要0.6美元

回到开头的可视化——那些跳动的线，就是Kimi K2.6在执行的信号。而指挥它们的，是Opus 4.8。这套架构的核心，就是把"大脑"和"手"分开。

第一部分：Opus做编排的5个提示词

Opus 4.8自带三个大部分团队没用到的功能：Dynamic Workflows管理上百个并行代理的任务流，Effort Controls自动分配算力，Extended Independence支持更长时间的自主运行。这些都不需要Opus自己干活，只需要它把计划做清楚。

提示词1：Opus做项目规划

I need to execute the following project using multiple AI agents:

Project: [描述你需要的最终结果]
Scale: [任务/文件/项目数量]
Quality standard: [合格输出是什么样]
Constraints: [截止时间、格式、限制]

Your job is to build the execution plan, not execute it.

Produce:
1. The breakdown of work into parallel streams
2. The dependencies between streams (what must happen before what)
3. The quality checkpoints where output must be reviewed before continuing
4. The exact brief for each execution stream (what each agent needs to know to do its part)
5. The final assembly instructions (how the outputs combine into the finished deliverable)

Do not write any of the actual content. Only the plan.

提示词2：动态工作流任务分类

I have the following tasks to complete this week:

[列出你的所有任务]

Sort them into two groups:

Group A — Opus handles (reasoning-heavy, judgment calls, ambiguous, high-stakes):
[Opus级标准：架构决策、质量审查、战略分析，任何出错都会带来后续问题的任务]

Group B — Kimi Agent Swarm handles (execution-heavy, batch processing, repetitive at scale, clear output spec):
[Kimi级标准：调研N个来源、生成N个文件、转换N个输入，任何输出标准清晰、可以并行的任务]

For each Group B task, write the exact project brief that Kimi Agent Swarm will receive.
For each Group A task, work through it directly.

提示词3：执行前定义质量标准

Kimi Agent Swarm is about to execute the following project:

[描述项目]

Before it runs, I need a quality rubric.

Define:
1. What a passing output looks like for each deliverable type
2. What a failing output looks like (be specific — not "low quality" but "missing citations," "under 500 words," "wrong format")
3. The 3 most common failure modes for this type of task
4. How to catch each failure mode in the output before it reaches the final assembly stage

This rubric will be used to review Agent Swarm outputs before they are accepted.

提示词4：Opus审查Kimi输出

Kimi Agent Swarm completed the following project:

Project brief: [项目要求]
Output: [粘贴或总结输出]

Review against the quality rubric:
[粘贴提示词3生成的质量标准]

For each deliverable:
- Pass or fail against each rubric criterion
- Specific issues that need to be fixed (quote the exact problem)
- Whether the issue requires a full re-run or a targeted fix

Produce a revision brief: exactly what Kimi needs to fix, in the same project brief format, so it can be sent directly without re-explanation.

提示词5：Opus整合Kimi输出得到最终结果

Kimi Agent Swarm produced the following outputs:

[列出或粘贴所有输出]

These are the components of: [描述最终交付物]

Your job is to assemble them into a coherent final output.

Rules:
- Do not rewrite what works. Connect and integrate.
- Identify and fix inconsistencies between sections
- Ensure the final output reads as one unified piece, not a compilation
- Flag anything that requires my review before finalizing

Final format: [准确描述组装后的输出应该是什么样]

第二部分：Kimi做执行的5个提示词

Kimi K2.6 Agent Swarm可以同时运行300个领域专用子代理，跨最多4000个步骤协调，输出可以直接生成PDF、表格、网站、数据集、代码这些真实文件。

算笔账：

• 调研50个竞品落地页：外包要2.5万美元，Kimi只要4-6美元的token费
• 批量写100封定制推广邮件：文案要2000-5000美元，Kimi一次就能跑完
• 30个代码库的技术审计：顾问要1.5万-4万美元，Kimi跑Opus定好的规则，最后Opus审核总结就行

提示词6：把Opus计划转成Kimi项目brief

Take the following execution plan produced by Claude Opus:

[粘贴提示词1生成的Opus计划]

Rewrite it as a Kimi Agent Swarm project brief.

Format:
Project: [一句话总结]
Input: [附加的文件、URL、数据]
Output: [文件类型/数量/命名规则/格式]
Phase 1: [第一个执行流 — 代理做什么，产出什么]
Phase 2: [第二个执行流 — 标注依赖关系]
Phase 3: [组装 — 输出怎么合并]
Quality note: [每个输出必须满足的最低标准]

The brief should be complete enough that Agent Swarm can execute without clarification.

提示词7：批量执行带输出规范

Project: [描述批量任务]
Input: [N个项目 — 附加文件或列出清单]
Output spec: [准确格式、文件类型、命名规则、每个输入对应一个输出]

For each input:
- [第一步要做什么]
- [第二步]
- [第三步]
- Output: [交付物准确格式]

Quality standard: [最低要求 — 字数、引用格式、结构等]

Run all [N] inputs in parallel. Deliver as [文件格式] named [命名规则].

提示词8：调研到交付一步完成

Research phase:
Search for [主题/竞品/对象] across [N] sources.
For each source extract: [列出具体数据点]
Output: structured dataset with one row per source.

Analysis phase:
Using the research dataset, identify: [patterns / gaps / opportunities / rankings]
Flag any source where the data is unclear or contradictory.

Deliverable phase:
Produce [最终输出格式] using the research and analysis.
Format: [准确规范]
Length: [字数或页数]
Citations: [格式]

Total output: one [文件类型], one supporting dataset.

提示词9：把工作流存成可复用技能

We just completed the following workflow:

Opus defined: [Opus做了什么计划]
Kimi executed: [Kimi跑了什么]
Output: [产出了什么]

Save this as a reusable Skill called [名称].

Document:
- The trigger (what kind of request activates this skill)
- The Opus orchestration prompt (what to send Opus to generate the brief)
- The Kimi execution brief template (what gets sent to Agent Swarm)
- The Opus review checklist (what Opus checks before accepting output)
- Expected inputs and outputs

Next time we run this workflow, I want to start from the Skill, not from scratch.

提示词10：工作流成本跟踪

This workflow just ran:

Opus 4.8 tasks:
[列出每个Opus处理的任务，估算token数]

Kimi K2.6 tasks:
[列出每个Kimi处理的任务，估算token数]

Calculate:
- Opus cost: [tokens] × $0.015 per 1K tokens
- Kimi cost: [tokens] × $0.0006 per 1K tokens
- Total actual cost
- What this workflow would have cost running entirely on Opus
- Savings this run

Log this to WORKFLOW_COSTS.md with date, workflow name, and breakdown.

第三部分：路由分类系统

从6.2万美元降到129美元，核心不是用Kimi代替Opus，是搭了一层路由分类：每个进来的任务先判断，需要判断还是需要执行，不同任务给不同模型。

路由规则很简单：如果能写出机器可评分的规则，就给Kimi执行；写不出来，就给Opus处理。

提示词11：每周工作流审计

Review the following workflows we ran this week:

[列出本周工作流]

For each workflow, classify every task:
- Opus-only: judgment-heavy, no clear rubric
- Kimi-only: execution-heavy, clear output spec, parallel-friendly
- Hybrid: Opus plans + Kimi executes + Opus reviews

For each Hybrid workflow, write the handoff points:
- What Opus produces before handoff (the brief)
- What Kimi receives (the execution spec)
- What Opus reviews after (the quality rubric)

Flag any task currently on Opus that should move to Kimi.

提示词12：生成路由决策树

I need a routing framework for our recurring workflow types.

For each workflow type below, define:
1. Which model handles each stage (Opus / Kimi / both)
2. The trigger that routes it to the right model
3. The handoff format between models
4. The cost estimate per run

Workflow types:
- [类型1，例如 内容调研+生产]
- [类型2，例如 竞品分析]
- [类型3，例如 推广信个性化]
- [类型4，例如 代码审查+重构]

Output: a routing table I can use to classify any incoming request in 30 seconds.

提示词13：月度成本优化审查

Last month we ran the following volume through our AI workflow:

[描述处理量 — 例如 200份调研brief，500封推广邮件，30次代码审计]

Current setup: [描述当前Opus和Kimi的分配]
Current monthly cost: [金额]

Analyze:
1. Which workflows are over-allocated to Opus that Kimi could handle
2. Which workflows are currently on Kimi that Opus should be handling for quality
3. What the optimal split would look like
4. Projected cost at the optimal split

Produce a routing change recommendation with estimated monthly savings.

提示词14：交接提示词模板

Create a standard handoff template for the following workflow:

When Opus completes [任务类型], it produces [输出格式].
Kimi Agent Swarm receives this and executes [执行任务].
Opus then reviews [审查标准].

Write:
1. The Opus output format that serves as the Kimi brief (structured, no ambiguity)
2. The Kimi execution brief template (slots for Opus to fill in)
3. The Opus review checklist (5 criteria Kimi's output is graded against)
4. The revision loop: if Kimi output fails, what gets sent back and in what format

This template becomes the permanent interface between the two models for this workflow type.

提示词15：给股东的ROI报告

Produce a monthly ROI report for our Opus 4.8 + Kimi K2.6 workflow setup.

Include:
1. Total workflows run this month: [N]
2. Volume processed: [N items / files / tasks]
3. Cost breakdown:
   - Opus 4.8 spend: $[amount]
   - Kimi K2.6 spend: $[amount]
   - Total: $[amount]
4. Equivalent cost running all workflows on Opus 4.8 alone: $[calculated amount]
5. Monthly saving: $[amount]
6. Annualized saving: $[amount × 12]
7. Quality incidents (workflows that failed review and required re-run): [N]
8. Quality incident rate: [%]

Format as an executive summary. One page. Numbers first.

最后

2026年5月28日之后，这套逻辑就通了：Anthropic做出来的Opus 4.8 Dynamic Workflows，就是用来管上百个并行代理的，不是让它当所有代理；Kimi K2.6的Agent Swarm，就是做大规模并行执行的，成本只要每百万tokens 0.6美元。

它们不是竞争对手，是同一个系统不同层的两个组件。想明白这点的团队，已经用每月129美元跑出了原来6.2万美元的效果，质量没降，速度没慢，只是价格回到了它本该有的水平。

当大家都动动嘴就能完成任务时，完成任务本身就不再是壁垒，成本控制就成了胜出的关键。