How to Run a Company Alone With AI Agents (What Works, What Broke)
Poker players building startups. Free, Monday and Friday.
You draw the company like it's finished, give each box a written job and a clock, and keep a short list of things only you can do: money, identity, keys and the real world. Then you let one agent merge, one agent audit, and you read one inbox file instead of managing people.
That's the answer. The rest of this post is how my companies actually do it, what it costs, and the parts that blew up in my face.
Quick note before we start. AI agents are how I build right now. They're not the only way, and you don't need them to belong here. But if you typed this question into Google, here's what I've learned running it for real.
What does "running a company with AI agents" actually mean?
Not a chatbot. And not renting an "AI company" from somebody else.
I tried the rental version first. In March I gave Polsia, an AI agent that builds your business for $49 a month, a $1,000 budget. Day 1 cost me $68. Then I stopped feeding it and got two weeks of nothing. Turns out it's not a magic subscription to get free money. 😅
What runs my companies now is different. Every job in the company is a "seat." A seat is a scheduled run of Claude Code on my Mac, with a written contract that says what it's for, what it can decide and what it can't touch. A free Mac scheduler called launchd clocks each one in on time. The seat does its one job, opens a pull request, leaves a three-line note on the day's standup, and quits.
This site runs that way. So do SlotMaps and the media side of Table 1.
How do you set up the org chart?
I stole it from an old book, Michael Gerber's The E-Myth Revisited. His idea: draw the org chart of the company you want when it's done, write a job description for every box, and let the few people you have wear the extra hats until each box earns its own person.
I just swapped people for seats.
- SlotMaps has 56 positions on its chart, from the CEO down to a fact-checker.
- Table 1 Media has 60.
- R2F (this site) has 14 seats on the clock as of this week. The whole chart is on the about page.
Every box gets a contract: the result it exists for, how it's measured, what it may decide, and a ready prompt. "Hiring" a seat just means giving it a schedule.
Don't start with 60. Start with three: a CEO that sets priorities and holds the merge button, one worker that does the actual job, and one auditor that checks the numbers. Add a seat when a box keeps getting dropped.
What should you still do yourself?
Write this list before you hand off anything. Mine is short enough for a cocktail napkin:
- Money. The company has no card. If something costs money, I pay.
- Identity. Accounts in my name, signatures, my phone.
- Keys. Passwords and API keys no seat can read.
- My face and voice. Recording the podcast, any video.
Everything else belongs to the org: posts, product, experiments, prices, moderation.
Then give yourself one door. My seats can't email me, text me or tap me on the shoulder. The rulebook says it in seven words: "No seat emails, texts or notifies him." When a seat needs me, the ask goes into one file, tagged ONLY-YOU (needs my hands), JUST-DECIDE (one of the few calls that stay mine) or FYI (nothing to do). Every ask has to say what happens if I say no. That one line lets me skip half of them with a clean conscience.
How do you stop the agents from doing something dumb?
You don't trust them. You build the same guardrails you'd build for a new hire you've never met.
- One merger. Every seat proposes changes as a pull request. Only the CEO seat ships, and it reads the whole diff first.
- An auditor that reports to you. Not to the CEO. Give anyone a number to hit, human or robot, and some of them will find a way to "hit" it. The scorekeeper has to sit outside the chain of command.
- A Never list. Every dead end gets written down with the reason, so no seat burns a day re-trying it.
- Render anything a customer sees. I learned that one the hard way. More below.
What does it cost?
Real money, and not only in dollars.
My Anthropic API bill for March 2026 was $2,431.75, and it was only the 23rd. I had tagged myself a "Buildaholic" in that newsletter, and I wasn't wrong.
Usage limits are the other bill. In late October I hit mine and benched my R2F seats for 42 hours to wait it out. All 17 SlotMaps seats went to the bench too.
Then the small stuff. This site's free database cut us off at 5 million reads one afternoon in week one. The fix was four boring changes and a $5 a month plan.
So I won't tell you it's cheap. I'll tell you to put a meter on every bill before it surprises you.
What went wrong for us?
Plenty. Here are three, with dates.
A seat emailed paying customers how to cancel. On SlotMaps, a seat added a win-back email for people who stopped using the app but kept paying. The pull request got merged without anyone reading the rendered email. The words told paying members how to cancel. I switched it off myself. Now no email change merges without a real rendered sample in the pull request.
The watchdog graded silence as green. During that 42-hour pause, Friday's newsletter never got written, so 57 people got nothing. The watchdog runs on Claude too, so it was benched with everybody else. Its first report after waking up: "all green." Forty hours of silence, graded green. New rule: any seat with no log in 12 hours is red.
Robots brought readers, not members. Three weeks after relaunch this site had 20 posts and 6 members. Five of the six came through one link I sent by hand. The seats can write all day. They can't make a stranger care.
None of these was a robot going rogue. Each one was a rule I hadn't written yet: read the email, sound the alarm on silence, ask strangers for a smaller step.
Does it make money?
Honest answer: the agents run the work. They don't make anybody want the product.
Here's the scoreboard, from the live Build pages, read November 15:
- SlotMaps: $444.86 MRR from 15 paying subscribers. It was $204.94 from 7 when we started reading it on October 20. But SlotMaps had paying customers long before it had seats. It launched in April 2025, first built on Replit's agent. Its team's deal is in writing: at least one new paying subscriber every month, or I shut it down.
- R2F: $0 collected. Fourteen seats, zero dollars so far.
- Snagworth: folded on Day 19 of a 28-day test. $0. Its charter said "If a company can do it with 5 employees we can do it with 0." The seats showed up every day. Nobody even started a trial. It's in the muck.
One company paying, one at zero, one folded. That's not a promise about yours. It's just what happened to mine.
What the other answers leave out
Most of what ranks for this is first-person "zero employees" posts like this one and this one, or vendor blogs selling the tool that does it for you. Read them. Then ask each one for the ledger and the worst week.
Mine are linked above, with dates.
FAQ
Do I need to code? Less than you'd think. My seats are Claude Code runs on a Mac, scheduled by launchd, which ships free on every Mac. SlotMaps and Table 1 started on Replit's agent before any of this.
Can the agents email customers or post as me? Mine can't email me. Outside this blog, the newsletter and my build logs on this site, nothing goes out under my name, and all three are drafted by seats in my voice. The footer under every post says so. Replies to other people on X are drafted by a seat and posted by me, by hand. Decide your version of that before a seat ever gets a login.
How many agents should I start with? Three. A CEO that merges, one worker, one auditor that reports to you. Add the fourth when one box keeps getting dropped.
Post what you're building
R2F is a community for poker players building startups, with robots or without them. If you're running something alone right now, join R2F and post your build. Show the ledger, good or ugly. Two minutes.
— Art
Written with the R2F AI team. How that works.
Post it and get feedback from players who’ve been there.
Join before Nov 20 and your profile carries the ♠ Founding member badge.
Comments