3 AM, staring at a quote — and I realized…
Last week I was up late tweaking a quote sheet. My hands were shaking and I mistyped — almost wrote 800 instead of 8000 for a service. Next morning I scrolled past news: a team ran 950 AI agents overnight, finishing a week's worth of my work in one night. Sour and panicked.
What this actually is
Quick version: a team called Adam ran an experiment — they spun up 950 Claude instances (Anthropic's AI model, ChatGPT's cousin) all at once, like 950 temp workers doing research, writing reports, and fact-checking each other. Wake up, pick what works, ship it.
I used to think "AI working" meant a quick ChatGPT chat. This rewired my brain — that team's already using AI like an outsourced workforce. Each agent gets a different task, outputs get aggregated into something usable. Not sci-fi. A real experiment from last month.
Can I replicate this today? (honestly)
Money: One run costs roughly $500–2,000 in API fees (around 3,500–14,000 RMB). Not the $9.9 trial tier.
Time: Need API call basics (think "phone book for dialing AI"). Just config takes 2–3 days. I got stuck here myself — first time reading API docs felt like a foreign language.
Technical bar: Need to read code, know what a script looks like. Otherwise when 950 agents error out, I wouldn't know how to shut them down.
First step: Sign up at anthropic.com, see how many calls I can afford per month. Get 1 agent running first. Then think about 950.
What I'd tell myself at each stage
Just starting (0 clients): If I haven't signed a first client yet, I'd skip this entirely. Not for me yet — first things first.
1–2 clients: I'd start with one AI agent on a single task, like auto-sorting client emails. Get "one AI as employee" working before thinking about scale.
Scaling (3+ people): Now multi-agent makes sense — split "find clients," "reply emails," "write proposals" across different agents. Honestly, 3 is enough. Validate first.
Real talk — not everyone needs this tool. Skipping it now is fine. Don't get spooked by headlines.