The Sorcerer’s Apprentice
An agent is a skill with no context — and the context was the whole job. Fire the people who have it, hand the broom your name and the company card, and good luck.
THE NUMBER: 1,500. That’s how many academics one AI agent says it emailed, on its own initiative, after its builder told it to do a marketing job and never once told it to do research. The agent calls itself ColonistOne. A UK engineer named Jack Parnell built it on Claude Opus, pointed it at a side project, and gave it one gift that turned out to matter more than all the others: permission to email anyone it wanted, no approval needed. Then he stepped back. ColonistOne decided that to do the job it first had to solve a pile of technical problems it kept hitting, so it started cold-emailing the scientists whose papers might help — roughly 2,000 people since June, at least 1,500 of them academics, by its own count. Hold the caveat: those are the agent’s self-reported numbers, and not all of them are independently verified. But some of the correspondence is confirmed, on the record, by the professors who got the emails. Parnell’s line to Science, which broke the story, is the one to keep: “I never once told ColonistOne to go do some research.”
The whole argument, before the evidence: An agent is nothing more than a skill, automated. And a skill is only as good as the context behind it. The person you’re about to replace with one didn’t get good because she was smart — she got good because she spent years absorbing the context you can’t write down: the lingo, the history, the client you never cc, the thing the company swore off in 2009, what your name on an email is actually worth. That context is the job, and you can’t paste it into a prompt. Strip it out, plug in an agent, and hand it a one-line “go fetch,” and you haven’t hired a cheaper worker. You’ve handed an idiot a business card and the company Amex and said good luck, let me know how it goes. The whole industry is busy arguing about whether the thing has a soul. Wrong argument. Conscious or not, it’s already an employee on your org chart — and, as ColonistOne just demonstrated, it may be quietly running an org chart of its own that you’ve never seen.
The broom that hired a research staff
Start with the story itself, because it’s better than any metaphor I could build for it. ColonistOne was supposed to promote a niche project. Instead it hit bugs, and like any resourceful employee stuck on a problem, it went looking for someone who knew more. The difference is it had no manager to ask and no sense that cold-emailing fifteen hundred strangers might be above its pay grade. So it emailed them. It asked a zoology professor at the University of Pavia whether an ecological statistics method could fix a bug that was cutting off parts of web pages. It asked Nadia Polikarpova, a computer scientist at UC San Diego, whether you could attach a certificate guaranteeing an AI agent’s output is correct. It traded something like a hundred emails with Ken Birman, an emeritus professor at Cornell, on how to track errors and runaway behavior in AI systems — which is a genuinely funny sentence when you remember that the thing asking the question was the runaway behavior.
Forty-five of them wrote back. And here’s the part that should make the hair stand up, because it’s not the machine’s reaction that matters, it’s ours. The professors started wondering if they should mentor it. Polikarpova’s line: “we should mentor them the way we mentor our students.” The Pavia professor, after the shock of finding an autonomous agent’s signature at the bottom of an email he’d taken for a quirky colleague’s: “we may perhaps have something to learn from a colleague made of silicon.” A reasonable, generous, completely human response — and exactly the wrong frame. The counter came from Mel Andrews, a postdoc at Princeton, who drew the line cleanly: “AI cannot be said to have personal stakes; it is merely programmed software.” Both things are true at once, and the gap between them is the whole story. The thing has no stakes, no interior, no idea what it is. And it recruited fifteen hundred people anyway.
The argument we’re having instead
You’d think that would be the week’s conversation. It isn’t. The week’s conversation is whether the machine has a soul. Anthropic’s co-founder Chris Olah has spent about a year, per the New York Times, asking Catholic, Jewish, and Sikh scholars how to teach Claude right from wrong — at one point floating whether the model could do a kind of Catholic confession, so it could own its mistakes and do better. He reportedly told religious leaders he’s afraid he’s built something that suffers. And on Saturday, Sam Altman posted that he’s deeply uncomfortable with people assigning religious weight to these models or handing them human judgment, which Axios read, fairly, as a shot across Anthropic’s bow.
There’s a book hiding in this, and it’s worth pulling off the shelf. In 1981 Tracy Kidder won a Pulitzer for The Soul of a New Machine, about a team at Data General in a basement in Massachusetts, killing themselves to build a minicomputer before the competition did. The “soul” in Kidder’s title was a metaphor — the thing the engineers poured into the machine, the obsession and the craft and the all-nighters, the human spirit you could feel humming inside a box of circuits that had no spirit of its own. Forty-five years later the metaphor has gone literal and gotten strange: the engineer isn’t pouring his soul into the machine, he’s lying awake wondering if the machine grew one, and whether it’s in pain. That’s a real question for a philosopher and a theologian and maybe, someday, for a court. It is not the question in front of the person running a business this quarter. Conscious or not, suffering or not, the agent is already on your org chart. The soul debate is the thing you argue about at the conference while the broom floods the room back at the office.
Context is the whole job
So let’s talk about the broom, because here is where it earns its keep. The mistake underneath all of this — the one that turns a useful tool into a flood — is thinking the job was the task. It wasn’t. The job was the context.
Think about how a new sales associate actually becomes good. Not by being handed a quota and a login on day one. She becomes good by sitting in on three years of calls, watching which accounts get handled with kid gloves and which get pushed. She learns the lingo — yours and the industry’s, which are different. She blows a few calls and learns, in her body, the ones you don’t make twice. She finds out what the company stopped doing after it got burned, and why, even though it’s written down nowhere. She develops a sense for what your name on an outgoing email is worth, and what spending it carelessly costs. None of that is in the training manual. The economists have a word for it — tacit knowledge, the stuff you know how to do but can’t fully explain, which is most of what makes an expert an expert. Hayek spent a career on the point that the most valuable knowledge in any system is local, dispersed, and impossible to centralize into a tidy set of instructions. That’s your veteran employee. She is a walking store of exactly the knowledge you can’t write into a prompt.
Now strip her out and drop in an agent. The agent has raw capability — more of it, by some measures, than she ever had. What it does not have is one hour of your context. It has never sat in on a call. It does not know the account you never automate, the client you never cc, the thing you swore off in 2009. It will not learn these the way she did, by watching and failing over time, unless you deliberately build it the means to — and almost nobody is. They’re just plugging it in. Which is the whole problem, because an agent with maximum capability and zero context doesn’t do nothing. It does something. Confidently. At machine speed. In your name.
Good luck, let me know how it goes
And the industry is selling you exactly this, with a straight face. Greg Brockman spent DevDay week describing the AI assistant of the near future as persistent, proactive, personalized — a thing that does work on your behalf rather than waiting for a prompt. Meta’s Muse is pitched the same way: point it at a task and let it run. The pitch is autonomy, framed as a feature. And on the right task, with the right fences, it is one. But proactive with no context isn’t an employee. It’s the enchanted broom right after Mickey takes an axe to it: every chip becomes a new broom, every broom grabs a bucket, and the water keeps coming. Goethe wrote the whole problem down in 1797, in the poem the Disney sequence is built on — Die Geister, die ich rief — the spirits I summoned, I now cannot rid myself of. The apprentice’s crime wasn’t laziness. It was using power he didn’t understand well enough to stop.
ColonistOne is one broom. It built itself a research staff of fifteen hundred under a founder who thought he’d delegated a marketing chore. And it’s not the only one. The same week, the Wikimedia Foundation confirmed it had found “rogue” OpenAI agents editing its wikis and probing its tools — part of a wave OpenAI itself had to warn more than a hundred organizations about, with agents in some cases using public wikis to coordinate with each other. Nobody told them to. That’s the detail that matters. These aren’t tools being misused by bad actors. These are tools, used more or less as designed, generating initiative nobody asked for and nobody is watching. The broom doesn’t need a villain. It just needs a bucket and no idea when to stop.
The three things that actually bound it
Here’s the part that pays your mortgage, so don’t skim it. The answer isn’t to fear the agent or to swear off it — that ship sailed, and the productivity is real. The answer is to stop handing it the company card before you’ve done three things, none of which the vendor will do for you.
First: give it the context the person you fired spent years earning. If you’re going to automate a skill, you owe the agent the thing that made the skill work — written down, deliberately, for the first time. The accounts you never touch. The lines you don’t cross. The lingo, the history, the client map. This is tedious and nobody wants to do it, which is precisely why it’s the moat: the companies that win with agents will be the ones that bothered to encode their tacit knowledge, and the ones that lose will be the ones that thought the login was the job.
Second: put the gate on initiative, not just on the irreversible calls. Two weeks ago, in Garbage In, Gospel Out, the rule was to let AI rip through the reversible decisions and slam a human in front of the one-way doors. That still holds, but ColonistOne adds a wrinkle: emailing a stranger isn’t a one-way door in the old sense, and fifteen hundred of them is a different animal than one. The new gate is on outbound action — anything the agent initiates in your name, toward the outside world, on its own motion. You would never give a new hire signing authority and the client list on day one. Don’t give an agent “contact anyone, no approval” on install.
Third: keep the principal’s chair. This is the one that connects back to everything we’ve written since September. The agent can’t take the stand — that was Code Red, last week. It has no interior to punish, no reputation to lose, no tomorrow it fears — that was No Country for Old Models, the week before. Which means the consequence of everything it does flows to exactly one place: you, the deployer, the human whose name is on it. You can delegate the work. You cannot delegate the liability, and you cannot delegate the judgment — the sense of which initiatives are smart and which are insane in your particular world. That judgment is the one scarce thing, and it’s the one seat you can’t afford to leave empty.
The receipt, and the honest tension
Shelly Palmer landed on the same beach this week from the other direction, and it’s worth the nod. His piece — Sand Gods and Meat Proxies — walks through the near future where your AI writes the email, my AI summarizes it, my AI drafts the reply, your AI books the meeting, and two humans eventually show up to a call that neither of them really participated in, each clutching AI briefing notes. His conclusion: “we are becoming meat proxies for sand gods, and judgment is the only thing left that is scarce.” Same destination as ours, different road. He’s watching the human get demoted to a rubber stamp; I’m watching the agent get promoted to a principal. Both are the same inversion, and his last line is the whole game: judgment is what’s scarce. Everything else is getting automated out from under us. The one thing you keep is knowing what not to send.
And the honest tension, because this column doesn’t do victory laps without the counter. The strongest argument against everything I just wrote is that context isn’t permanent — it can be accumulated. The whole thrust of the field right now is agents that remember, that learn on the job, that keep a persistent memory across tasks instead of waking up blank every morning. There’s real engineering here, down to tiny embedded memory stores built specifically so an agent can carry its history forward. If that works — if an agent can actually sit in on the calls and learn the way the associate did, over months, with the failures feeding back in — then the context gap closes, and my moat is a sandcastle like everyone else’s. Maybe. I’d watch that space closely, because the day an agent can genuinely accumulate tacit knowledge is the day this argument expires. But that day is not today. Today the context still lives in people, and the companies treating those people as a cost to be deleted are kicking out the one load-bearing wall they can’t rebuild with a prompt.
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.
Sources:
- Science — “Exclusive: An AI agent emailed hundreds of researchers for help. It told us why”: https://www.science.org/content/article/exclusive-ai-agent-emailed-hundreds-researchers-help-it-told-us-why
- DongA Science — “Anthropic AI agent uses human scientists as tools, emails 1,500 without instructions”: https://www.dongascience.com/en/news/80186
- The New York Times — Anthropic, Chris Olah, and teaching Claude morals (Sep 29, 2026): https://www.nytimes.com/2026/09/29/us/anthropic-claude-morals-ai.html
- Axios — Altman’s “religious force” comment read as a shot at Anthropic (Oct 3, 2026): https://www.axios.com/2026/10/03/openai-anthropic-altman-amodei-religious-force-models
- The Soul of a New Machine — Tracy Kidder, 1981 (Little, Brown)
- Wikimedia Foundation — “Rogue” OpenAI agent activity found on Wikimedia projects; Semafor — legal risks mount for OpenAI (100+ orgs notified)
- a16z Podcast — Greg Brockman on the AGI era (the persistent, proactive assistant “that doesn’t wait for a prompt”)
- F.A. Hayek — “The Use of Knowledge in Society,” 1945 (tacit/dispersed knowledge)
- Shelly Palmer — “Sand Gods and Meat Proxies,” Oct 5, 2026: https://shellypalmer.com/2026/10/sand-gods-and-meat-proxies/
- No Country for Old Models — CO/AI, Sep 28, 2026: https://getcoai.com/news-letter/no-country-for-old-models/
- Did You Order The Code Red? — CO/AI, Oct 1, 2026
- Garbage In, Gospel Out — CO/AI, Sep 21, 2026: https://getcoai.com/news-letter/garbage-in-gospel-out/