300 个智能体，一个提示词，输出真实文件：Kimi 的隐藏利器

300 agents. One prompt. Real files as output. Most Kimi users have never opened Agent Swarm.

100 job descriptions turned into 100 tailored CVs. One prompt.

40 academic papers turned into a 100,000-word literature review with full citations. One prompt.

30 brick-and-mortar stores without websites, found on Google Maps, each given their own landing page. One prompt.

This isn't a plugin. This isn't a script. This is one button in the left sidebar of Kimi that most users have never clicked.

It's called Agent Swarm.

Every time you open Kimi, you get a chat window.

Type a question. Get an answer. Close the tab.

That's how 90% of people use it. It works. But it's 10% of what the product actually does.

Agent Swarm is not a chat interface. It's a multi-agent orchestration system. You give it a project brief. It spins up to 300 domain-specialized sub-agents. They work in parallel. Each one handles a piece. They coordinate across up to 4,000 steps. Output lands as actual files: PDFs, spreadsheets, websites, datasets, code.

The difference between using Kimi as a chatbot and using Agent Swarm is the difference between asking someone to research a topic and hiring a 300-person research firm to deliver a finished report.

Here's the financial reality of what Agent Swarm replaces.

A professional CV writer charges $150-300 per resume. 100 tailored CVs for a job hunt: $15,000-$30,000. Agent Swarm: one prompt. One sitting.

A research assistant at $50/hour takes 200+ hours to synthesize 40 papers into a 100,000-word literature review. That's $10,000+ and four weeks of calendar. Agent Swarm: one prompt. Cited. Formatted. Done.

A web developer charges $500-2,000 per landing page. 30 landing pages for 30 local businesses: $15,000-$60,000. Agent Swarm finds the stores, builds the pages, and delivers them. One prompt.

Three tasks. $40,000-$100,000 worth of professional labor. One tool.

Professional CV service (100 CVs): $15,000-$30,000 Research assistant (100,000-word review): $10,000+ Web developer (30 landing pages): $15,000-$60,000 Agent Swarm: handles all three

→ Rule 1: Think in deliverables, not answers.

Agent Swarm is not a question-answering tool.
It's a project execution engine.

The right question: "What files do I need at the end?"
Not: "Can you explain X?"

Wrong: "What are the best landing page elements for local businesses?"
Right: "Build 30 landing pages for the 30 stores listed in this file.
       Output: one HTML file per store, named by business name."

→ Rule 2: Batch is where the leverage lives.

Agent Swarm is 10x more powerful when you have N of something.

100 job descriptions → 100 tailored CVs
40 research papers → full literature synthesis
30 business profiles → 30 websites
10 brief prompts → 10 magazine covers

If your task has a number in it, Agent Swarm is the right tool.

→ Rule 3: Four task types where it dominates.

1. Deep and wide research
   Multiple sources, multiple angles, synthesized output.
   Example: "Analyze all 40 attached papers. Identify consensus,
   conflicts, and research gaps. Output: 100,000-word review + cited dataset."

2. Large file batches
   Process many files in parallel instead of one at a time.
   Example: "Process each of the 100 attached CVs against each job description."

3. Multi-part analysis
   Tasks that require research then writing then formatting.
   Example: "Research 30 companies. Write a competitive brief for each.
   Format as a PDF deck."

4. Output-heavy work with real deliverables
   When the result needs to be a file, not a message.
   Example: "1 astrophysics paper → 40-page report + 20,000-row dataset +
   14 astronomy-grade charts, saved as a reusable Skill."

→ Rule 4: Skills multiply your output.

After Agent Swarm completes a workflow you'll run again:

"Save this entire workflow as a reusable Skill called [name].
Include: input format, agent steps, output format.
I want to run this same process next month with new files."

Your best workflows become one-click repeatable.

→ Rule 5: Real data beats abstract descriptions.

Agent Swarm runs on context.

PDF files beat summaries.
URLs beat descriptions.
Spreadsheets beat bullet points.

The more specific your input, the more specific your output.
"Analyze these 40 papers" + actual attached PDFs = full literature review.
"Analyze research papers about X" = generic overview.

Always attach the source files.

Agent Swarm is not the default Kimi chat. You have to switch to it intentionally.

Open kimi.com. Look at the left sidebar. You'll see: New Chat, Slides, Websites, Docs, Deep Research, Sheets, Agent Swarm, Kimi Code, Kimi Claw. Click Agent Swarm.

That's the room where the 300 agents live.

→ Rule 6: Write a project brief, not a question.

The wrong format:
"Can you help me analyze these job descriptions and create CVs?"

The right format:
"Project: Job application batch
Input: 100 job description files attached.
My background: [paste your work history, skills, 200-300 words]
Output: 100 tailored CV PDFs, one per job description.
Naming format: [Company]_[Role]_CV.pdf
Tone: professional, ATS-optimized, under 2 pages each."

The more specific the brief, the better the agent plan.

→ Rule 7: Specify output format before anything else.

Lead every Agent Swarm prompt with:
"Output: [file type] / [count] / [naming convention] / [format details]"

Examples:
- "Output: 30 HTML files, one per store, named by business name"
- "Output: 1 Word document, 100,000 words, Chicago citation style"
- "Output: 40-page PDF report + one CSV dataset with 20,000 rows + 14 PNG charts"

If you don't specify the output first, the agent decides.
Sometimes that's fine. Usually it's not what you wanted.

→ Rule 8: Feed it files, not descriptions.

Attach everything relevant before running the prompt.

Job hunting batch: attach all 100 job description PDFs.
Literature review: attach all 40 research papers.
Competitive analysis: attach competitor websites as URLs or scraped files.
Market research: attach your data CSV, not a summary of it.

Agent Swarm reads files the same way 300 humans would.
The more context per agent, the better the output per file.

→ Rule 9: Review the agent plan before execution.

After you submit a project brief, Agent Swarm shows you
the execution plan before it starts running.

Read it. Check:
- Does it understand the scope correctly?
- Is the step count reasonable for the task size?
- Does the output plan match what you actually need?

You can adjust here before it runs 4,000 steps in the wrong direction.
This is the most important step most first-time users skip.

→ Rule 10: Save repeatable workflows as Skills.

After completing any workflow you'll run again:

"Save this entire process as a Skill called [name]:
- Input format: [describe]
- Steps: [what the agents did]
- Output format: [what it produced]

Next time I run this, I want to attach new files and get the same output."

Skills turn one-time workflows into permanent infrastructure.

mistake to avoid: Don't write the project brief inside a regular Kimi chat and then paste it into Agent Swarm. Write directly inside Agent Swarm from the start. The interface gives you space to structure your brief properly.

The gap between a weak Agent Swarm run and a strong one is almost always in the prompt.

A chatbot prompt asks for information. An Agent Swarm prompt commissions a project. The mindset shift changes everything: you're not asking a question, you're briefing a team.

→ Rule 11: Open with the number.

Every Agent Swarm prompt should start with how many outputs you need.

"I need 100 CVs."
"I need 30 landing pages."
"I need a 100,000-word document."
"I need 14 charts."

The number anchors the swarm before it reads anything else.
Agents calibrate scope, resources, and step allocation from that number.

→ Rule 12: Break complex work into explicit phases.

For multi-stage projects, write the phases directly into the prompt:

"Phase 1: Research
 Search Google Maps for all retail locations near [city].
 Identify 30 stores without official websites.
 Output: spreadsheet with store name, address, category, contact.

 Phase 2: Build
 For each of the 30 stores: create a landing page.
 Include: business name, address, hours, description, contact form.
 Output: 30 HTML files, named by business name."

The agents follow your phases in order.
Without phases, they make their own plan. Sometimes that's fine.

→ Rule 13: Quantify the output inside the output spec.

Don't just say "a comprehensive report."
Say: "a 40-page report with executive summary, methodology section,
findings broken into 5 chapters of 6 pages each, and bibliography."

Don't just say "charts."
Say: "14 astronomy-grade charts: 5 scatter plots, 4 bar charts,
3 line graphs, 2 heat maps. All labeled, sourced, export-ready as PNG."

Specificity at the output level gives agents a quality target to hit.
Vagueness gives them permission to stop early.

→ Rule 14: Use real data as your anchor.

The single most effective thing you can add to any Agent Swarm prompt
is a real file that defines the scope.

100 job descriptions: attach the actual 100 PDFs.
40 research papers: attach the actual papers, not a reading list.
Competitive analysis: attach real competitor URLs, not company names.

One attached file does more work than three paragraphs of description.

→ Rule 15: End with the Skill instruction.

For any workflow you might run again, add this at the end of your brief:

"After completing this project, save the entire workflow
as a reusable Skill. Name it: [descriptive name].
Document the input format, agent steps, and output format.
I want to reuse this next month with new source files."

This turns a one-time run into a permanent system you own.
The first run builds the workflow. Every run after that is free leverage.

Here's what 300 agents actually replace.

100 tailored CVs: $15,000-$30,000 from a professional service 100,000-word literature review: $10,000+ from a research assistant 30 custom landing pages: $15,000-$60,000 from a developer Total manual cost: $40,000-$100,000

Agent Swarm: one prompt per project

The model underneath is Kimi K2.6. It scored 58.6% on SWE-Bench Pro. Claude Opus 4.6 scored 53.4%. It's open-weight. It costs $0.60 per million input tokens. Claude Opus costs $5.00.

The infrastructure is better than what most teams are paying for. The Agent Swarm layer on top of it does the parallel execution that no individual model can do alone.

15 rules. One sidebar button most users have never clicked.

The developers who find Agent Swarm stop doing one-at-a-time work. The ones who haven't are still asking one question at a time to a chatbot that can run 300 agents in parallel.

the most underused feature in Agent Swarm is Skills. the first run takes 20 minutes. every run after that takes 30 seconds. build the workflow once and you've built infrastructure. that's the actual leverage.