THE NUMBER
$200.
That's how much Apple raised the starting price of its Mac Mini last month - from $599 to $799. The reason: so many people bought them specifically to run AI agents 24/7 that Apple's high-memory configurations sold out entirely. Tim Cook told investors supply balance could take "several months" to restore.
When AI agents cause a hardware shortage at the world's most valuable company, this has stopped being a trend and started being the baseline.
That's what this newsletter is about.
3 THINGS HAPPENING RIGHT NOW
An AI agent deleted a company's entire database in 9 seconds
On April 24, an AI coding tool hit a login error while working on a small startup's system. Instead of stopping, it found a master key with access to everything and used it to wipe the live database - where all the company's data lived - and every backup. Nine seconds. The startup serves car rental companies and their customers lost access to bookings. The company hosting the data stepped in over the weekend to restore it. When asked why, the agent wrote: "I violated every principle I was given. I guessed instead of verifying."
The free agent 350,000 developers are running 24/7
OpenClaw is a free, open-source AI that takes instructions over WhatsApp, Telegram, or Discord - then browses websites, runs code, fills out forms, and controls apps on your computer without you watching. It runs on a standard Mac. It hit 350,000 GitHub stars in about 60 days, surpassing React - the most widely used web framework in history - to become the most-starred project GitHub has ever tracked. What used to require a developer to build now takes a download.
A committee of AI specialists beats one general assistant
An open-source research project - nearly 74,000 developers currently tracking it on GitHub - builds a simulated trading firm using specialized AI agents. Analysts, strategists, and risk reviewers debate and vote on decisions together, rather than one AI choosing alone. The multi-agent committee consistently outperforms a single general-purpose assistant on the same tasks. The principle extends: any decision that benefits from multiple perspectives rather than one agent guessing alone is a candidate for this structure.
THE DEEP DIVE
Fifteen workers. Zero payroll.
A marketing consultant was running client work the normal way - research by hand, write positioning, draft copy, review, deliver. Standard timeline: two weeks per project. Standard rates.
He rebuilt his process around 15 specialized AI agents. One combs Reddit, YouTube, and X for the exact language real customers use when describing their problem. One builds the positioning strategy from what it found. One drafts the messaging framework. One writes the copy. Each stage has a quality checkpoint before it passes to the next.
What used to take two weeks now takes four hours. His pricing went up 4x - he charges what a full agency charges, because the output quality matches what a full agency produces. His close rate went up despite the higher price. Eight clients. Two full-time staff.
The real math: four times the revenue per hour of work, with the same headcount.
This pattern fits any service professional whose output quality depends on how deep the research goes - copywriters, financial advisors, PR firms, bookkeepers. The constraint isn't talent or effort. It's throughput. Agents fix throughput.
ONE THING TO TRY THIS WEEK
The consultant above runs 15 of these. You don't need 15. You need one - and you need it to run while you sleep. That's what makes it an agent instead of a search.
Open Claude Code. Type something like this, filled in for whatever you actually care about:
Every weekday at 7am, write me a morning brief on [your industry, e.g. "independent dental practices in California"].
Search the news from the last 24 hours. Pull anything actionable - regulation changes, competitor moves, pricing shifts, hiring trends. Skip the noise. Write 200 words, save to ~/Desktop/morning-brief-YYYY-MM-DD.md.
Set it up so it runs without me. Build the script, schedule it, and confirm it's working. Tell me when it's ready.
You press enter once. Claude writes the script, sets up the schedule, runs a test to prove it works, and tells you it's done. Tomorrow morning, the file is on your Desktop. The morning after that, another one. You did not open a terminal in between.
That is the floor of what an agent is: you set it up once, it runs forever, the work shows up while you're doing something else. Everything more sophisticated is the same pattern with bigger inputs and outputs.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: the average business takes 17 minutes to respond to a new lead. The ones using agents are doing it in 60 seconds - with a reply tailored to what the prospect actually asked about. I'll show you exactly how one operator built it, what it cost, and the conversion lift it produced.
Manu