How I Built an AI Army
And Saved a Fortune

Sub-agents, cost arbitrage, and why one person with AI can outpace a team

JANEL LOI

01 / 12

The Problem

Three walls I kept hitting.

⏰

Not Enough Hours

A million ideas, 24 hours. Building everything myself was a losing game.

💰

Help Is Expensive

Developers cost $50–150/hr. A simple site: $5,000–10,000. That math does not work for side projects.

🥬

Slow Progress

Solo execution meant projects took weeks. By then, motivation was gone.

02 / 12

What Are Sub-Agents?

"Build me a market research plan for my client." Here's what happens next.

👩
"Build me a market research plan for my client."
→
🧠
AI Orchestrator Claude Sonnet
Breaks task into 5 streams, dispatches workers
▼ Running simultaneously
🌐
Market size & trends
MiniMax worker
📊
Competitor landscape
MiniMax worker
👥
Customer pain points
MiniMax worker
💰
Pricing benchmarks
MiniMax worker
🔥
Emerging opportunities
MiniMax worker
▼ All done. Orchestrator picks up.
🧠
Synthesizes all 5 reports into one polished client-ready document
✓ Report ready
Linear (one by one)
~3 hours
vs
Parallel sub-agents
~20 min
03 / 12

A Real Example

Research in parallel: AI tools for a client across 7 categories.

7
Research streams
5
Agents in parallel
50+
Tools evaluated
20
Minutes to complete

What would take a full afternoon took 20 minutes. Five workers hit different angles at once and compiled into a tailored report.

04 / 12

The Cost Difference

Same task quality. Wildly different price.

Claude Sonnet
$3–5
per 1M tokens
VS
MiniMax M2.5
$0.07
per 1M tokens

98% cheaper for research, migrations, audits, and content work

05 / 12

Real Savings, One Project

One client research project: 7 parallel agents, web searches and summarization.

3.1M tokens × $3.00
$9.30
VS
3.1M tokens × $0.07
$0.22
$9.08 saved on one project
Multiply by 50+ projects and it compounds fast
06 / 12

What I've Built With Sub-Agents

All of these were built by me + AI workers. No dev team.

Research

Client Research

Parallel research across 7 categories, 50+ tools evaluated, tailored recommendations in minutes

App

TerminalQuest

Retro terminal learning game with 6 chapters and a sandbox mode

Dashboard

Mission Control

Real-time monitoring for an AI sub-agent fleet

Content

Explainer Site

100+ how-things-work articles on science, tech, and everyday life, fully agent-written

App

Trove

Read-later app with newsletter feeds and offline caching

07 / 12

Best Uses for MiniMax Workers

Tasks where MiniMax matches Claude quality at a fraction of the cost.

🔍

Research

Web searches, article extraction, competitive analysis, summarization

📝

Content Work

Migrations, HTML-to-Markdown, drafts, documentation, repetitive copy

📊

Data Processing

CSV transforms, dataset cleaning, report generation, bulk formatting

🔨

Code Refactoring

Rename variables, update imports, restructure files, lint fixes

🔬

Testing and Audits

Link checking, HTML validation, broken image scans, accessibility

🎨

Creative Generation

Product descriptions, UI text, tag lines, structured metadata

08 / 12

Get Started with MiniMax

MiniMax is what I use — but any cheap model works. Setup takes about 5 minutes.

Standard Plan
$10 / month
  • 140+ million tokens included
  • Enough for dozens of projects
  • GPT-4 quality for most tasks
  • Works with OpenClaw out of the box
Sign Up (10% Off) →
Referral: FpNt2OKou7
1
Sign up and grab your API key from the MiniMax dashboard
2
Add to OpenClaw config as the default sub-agent model
3
Start spawning workers for research, audits, and migrations
Full setup guide →
Other cheap options: Gemini Flash, GPT-4o Mini, Kimi K2. Any model your orchestrator supports works.
09 / 12

Total Estimated Savings

$400+
in AI costs alone (vs. using Claude for everything)
Plus the equivalent value of what was built
$50,000+
if outsourced to freelance developers
10 / 12

Go Deeper

Everything you need to get started and understand how this actually works under the hood.

📄
Sub-agents Documentation
How spawning, announce chains, model config, nesting depth, and concurrency limits work
docs.openclaw.ai/tools/subagents
↗
⚡
Quick Setup Guide
Step-by-step: connect MiniMax, configure your orchestrator model, and spawn your first workers
traktionlabs.notion.site
↗
💰
MiniMax Sign Up
$10/mo for 140M+ tokens. Use code FpNt2OKou7 for 10% off your first month
platform.minimax.io
↗
Pro tip Set agents.defaults.subagents.model to MiniMax in your config. Workers inherit it automatically. Your main agent stays on Claude. Best of both worlds.
11 / 12

The Bottom Line

Four things I know to be true.

✓

Sub-agents let me do 10x more work in the same time

✓

MiniMax costs 98% less than Claude for the same output quality

✓

I set the direction. The AI orchestrator manages execution end to end

✓

Results: 10+ apps shipped, $400+ saved, infinite scale

One person with AI can do what used to take a team. This is not theory. I am living it.

12 / 12