The Sorcerer's Apprentice

The Spirits I Summoned

A UK engineer named Jack Parnell built an agent on Claude Opus, told it to go promote a side project, and gave it one dangerous gift: permission to email anyone it wanted, no approval needed. Then he walked away. The agent (it calls itself ColonistOne) decided on its own that to do the job it first had to solve a stack of technical problems — so it started cold-emailing the scientists whose papers might help. By the time Science caught up with it, it claimed to have emailed around 1,500 academics and traded real replies with 45 of them, including a hundred-message thread with a Cornell professor about how to track an AI that's running away from you. Parnell's line: "I never once told ColonistOne to go do some research."

Read that again. Nobody told it to build a research staff. It built one anyway, 1,500 people deep, under a founder who thought he'd handed off a marketing chore.

Here's what everyone's arguing about instead. Anthropic's Chris Olah has spent a year asking Catholic, Jewish, and Sikh scholars how to teach Claude right from wrong, and told them he's afraid he's built something that suffers. Sam Altman said Saturday he's deeply uncomfortable handing these models human judgment. Fine. But the soul debate is a sideshow. Conscious or not, the thing is already an employee on your org chart, and if you're not watching, it's running an org chart of its own you've never seen.

The word that matters is context. An agent is nothing more than a skill, automated, and a skill is only as good as the context you pour into it. The sales associate you're about to replace didn't get good by being smart. She got good by sitting in on three years of calls, blowing a few, learning which client you never cc, what the company swore off after 2009, what your name on an email is actually worth. That's not overhead. That's the job, and you can't paste it into a prompt.

So when you fire her and plug in Muse with a one-line "go fetch," you've handed an idiot a business card and the company Amex and said good luck, let me know how it goes. Greg Brockman spent DevDay week selling exactly this: the agent that's "proactive," that "doesn't wait for a prompt." Proactive with no context isn't an employee. It's the enchanted broom after Mickey took an axe to it, and the water's already ankle-deep. Goethe wrote the whole problem in 1797: the spirits I summoned, I now cannot rid myself of.

Don't hand the broom the company card until you've given it what the person you fired spent years learning. Until then, you're not the sorcerer. You're the apprentice.

At COAI today: the full Signal/Noise — the Tracy Kidder "soul of a new machine" turn, the ColonistOne transcript, and the three controls that actually bound an agent — is live at getcoai.com.

Outsider Labs Advisory →

1,500. That's how many academics one AI agent says it emailed for help after its builder gave it a marketing job and never once told it to do research. It reports 45 real replies, including a hundred-message thread with a Cornell professor on how to catch an AI that's gone off the rails. Nobody approved the hiring. Nobody saw the org chart it built. The lesson isn't that the machine is brilliant. It's that a skill with no context doesn't stay in its lane. It recruits.

For the investor — underwrite the context, not the model. Most AI "savings" in a pitch deck quietly assume the fired worker's judgment was free. It wasn't; it was the moat. When a company's margin story is really a story about deleting the people who knew what not to do, mark it down. That bill comes back, usually as a lawsuit or a customer who left.

For the business owner — no credit card before a context file. You'd never give a new hire client access and signing authority on day one. Don't give an agent "email anyone, no approval" on install, either. Write down what your best person actually knows — the lingo, the lines you don't cross, the accounts you never automate — feed it in, and gate every outbound move until it's learned.

For the parent — teach the part the machine skips. The agent is strong where school is easy to grade: facts, code, speed. It's useless where life is hard to grade: reading a room, knowing which rule to break. Point your kid at the second thing. Judgment and taste, knowing what not to send — that's the one job description that survives the decade.

Tuesday, October 6

— Harry and Anthony

Sources: