I Ran My Own Multi-Agent Swarm in 60 Seconds

Grok 4.20 just shipped a multi-agent swarm. Every query routes through 4 parallel AI agents that debate each other before answering. “Heavy” mode scales to 16. xAI says it cut hallucinations by 65%.

I thought: I can do that right now in my terminal.

The Prompt

One sentence to Claude Code:

Spin up 8 agents. Each one researches a different way to make $100 online in one week. Each picks a creative name, does its own web research, and saves a JSON proposal.

All 8 launched in parallel. About 60 seconds later, they filed these:

AGENT NAME               STRATEGY

The Ink Slinger          Freelance blog posts at $25-35 each

The Click Hustler        Stack UserTesting + Prolific + Respondent

The Thrift Alchemist     Thrift store flips on eBay/FB Marketplace

The Template Alchemist   Canva template bundles on Etsy

The Prompt Mechanic      AI-powered landing page builds

The Hundred Dollar Tutor Online tutoring at $25-35/hr

The Inbox Whisperer      Social media mgmt for small businesses

The Golden Mic           Podcast editing at $25-40/episode

The Blind Test

I picked the one I thought was most realistic. Then asked Claude for its pick, without revealing mine.

Same choice. The Click Hustler.

Claude’s reasoning: “Every other proposal depends on convincing a stranger to hire you — a brand new seller with zero reviews. The Click Hustler skips that entirely.” No client acquisition, earning starts within hours, highest floor.

The strategy: sign up for UserTesting, Prolific, and Respondent, then spend the week sniping high-paying tasks via notifications. No portfolio, no cold outreach. Just show up and get paid. Boring, but most likely to actually work.

The Point

Multi-agent isn’t a product feature. It’s a pattern. You can wire it up yourself with whatever model you prefer, right now, in a terminal. For a real application of this at scale, the AI influencer pipeline runs 120 parallel image generation jobs using the same fan-out architecture.

But the interesting part wasn’t the swarm. It was the blind test. I formed my opinion before asking the AI for its take. Same pick, same reasoning. That’s worth more than 16 agents agreeing with each other.

Next time you really want an answer to something, try it. Spin up multiple agents, let them work independently, then make your own call before asking the AI for its. It’s the best cure for the yes-man problem I’ve found.