I run my one-person business with ~17 narrow AI agents

A sole-proprietor developer and computer-science consultant has built a fleet of roughly 17 narrow AI agents, each scoped to a single job with its own memory and context, to run a one-person business spanning Amazon KDP puzzle-book publishing, crypto market-making bots, and bonus-hunting. In one documented case, a monitoring agent pulled trading logs at 3 AM, bucketed a $1.14 loss by cause, and traced it to a bot firing market orders that bypassed its own bookkeeping, then produced a fix brief with a dry-run plan, pass/fail criteria for a 24-hour live test, and a rollback plan. The author frames the setup as a distributed-systems design and is publishing the results as a build-in-public series.

One night earlier this month, my trading bot lost $1.14. Not a typo. One dollar and fourteen cents. It started the session at $4.85 and ended at $3.71. I know exactly why, because one of my AI assistants pulled the logs at 3 AM, split the loss into buckets price moves, maker rebates, fees, and a mystery “gap” , and found the culprit: the bot was firing market orders that bypassed its own bookkeeping. It rotated one coin in and out 58 times in five and a half hours, while that coin went up 31%. Then it wrote a change brief for the fix, with a dry-run plan, pass/fail criteria for a 24-hour live test, and a rollback plan. That’s a small, slightly embarrassing story, and it’s the best way I can explain what this newsletter is about. I’m a sole proprietor. I don’t have employees. What I do have is a fleet of AI assistants, each with its own job, and together they let one person run more experiments than he has any right to. This is the first post in a build-in-public series. I’ll show you what I’m building, the real numbers including the $3.71 ones , what flopped, and what you can steal. Why I did this My background is computer science: networks and distributed systems. I’ve done software development and consulting for years. I started using computers as a young kid, and it didn’t take long before I wanted to learn how to program them. That curiosity never really went away. The problem with being a one-person business is that you’re also the bookkeeper, the researcher, the marketer, the tax guy, and the person who should have read the terms and conditions. I had more ideas than hours. Then an opportunity to try AI agents came along, so I figured I’d jump in and see what happens. This newsletter is the record of what happened. So I treated it like a distributed systems problem. Instead of one giant chat where I ask everything, I split the work into separate agents, each with a narrow job, its own memory, and its own context. One agent knows everything about my puzzle books and nothing about my taxes. That separation turned out to matter more than I expected more on that below . The roster Here’s the team, grouped by what they do. Publishing - Books. Builds and maintains my Amazon KDP puzzle books. The source, print files, and listing drafts live in a git repo. Two books are live on Amazon: The Thief Stayed the Night a snowbound-hotel mystery puzzle book with 12 elimination cases and Frostwood Express . Both are in a cozy winter series I call Frostwood. Twelve more are built and print-ready but not published yet: logic puzzles, cryptograms, crosswords, a sketchbook, a couple of audio-first Kindle ebooks, and an advent puzzle book. Trading and markets - Trading strats and trade bot status. I run a set of small crypto market-making bots I wrote. There are four bots on one shared core, including a “ladder” bot and Avellaneda-Stoikov market makers on Coinbase and Kraken. The code is public at github.com/blakelapierre/crypto-mm-public. One agent helps design strategy, and the other watches the live bots, diagnoses losses, and writes up fixes that’s the 3 AM report from the intro . - Kraken prop trading. I trade a Kraken prop account by hand, since Kraken Prop has no trading API. The agent built me an alert-only breakout scanner. Every hour it checks the 67 perpetual markets I can trade on 1-hour candles, then prints entry, stop, and size. It also paper-trades its own signals so I can see whether the setup actually works. It never places orders. I do. - Options scanner and crypto. These scan for setups and data I care about, like cheap leveraged-ETF options and CBOE data pulls. The options scanner hunts for the cheapest contracts possible with outsized potential payoffs: tiny bets, asymmetric upside. Money hunting - Bonus harvesting , referral bonuses , and money making. These agents scan bank sign-up bonuses, brokerage promos, and referral programs. They track deadlines, requirements, and whether each offer is open to someone in California. The referral agent’s latest scan tracks 36 programs, values points conservatively at 1¢, and leads with an “URGENT: deadlines in the next few days” section. Another one researched trading competitions and incentive programs open to US residents, like Kalshi’s liquidity incentives. Running the business - Sole proprietor Odoo bot. Keeps my Odoo open-source ERP in shape: timesheets, reports, and the boring admin a real business needs. - US tax code. Answers tax questions from the actual statute. When uscode.house.gov was down for maintenance, it pulled the official Title 26 XML release from an Internet Archive capture of that exact URL instead of guessing. I love this kind of paranoia in a tax assistant. - Small business capital. Researched funding and business-credit options for a sole prop, plus non-dilutive grants like NSF SBIR, and tagged every claim as verified, third-party, unverified, or estimate. - SHC consulting. It listens to a channel in Odoo the open-source business software I use and does whatever’s needed when something comes in. Honestly, I’m a little fuzzy on the details myself, which says something about how much these things run on their own. Side projects - Streaming audio. A 24/7 YouTube live stream of generative music in 17 styles, synthesized in real time with numpy no samples, no copyrighted material and streamed with ffmpeg. Viewers vote on the next style, mood, and instrument from live chat, and a hype meter follows how busy chat is. - Web radio. Hardcore , a static web-radio site hardcoreradio.website that plays YouTube videos through the official player API, with thumbs-up/down ratings. Generalists - Overview and general questions. One keeps the big picture across everything, and one answers whatever doesn’t belong anywhere else. And yes, there’s now one helping me write this newsletter. How it works day to day Each agent is a Grok Bot assistant with its own memory and instructions. They all share one Linux box, so files one agent makes are available to the others. The book repo, the trading notes, and the scan outputs all live there. When the bonus agent finds an offer and the tax agent needs to know how that income is treated, the data is already sitting on disk. A typical week looks something like this: I check the bot chats, read what they’ve done, and send over any new ideas. That’s it. It takes exactly as long as it takes me to read and type. Some things run on a schedule. The money-scan agent runs hourly and only reports what’s new or materially changed since its last scan, so I don’t drown in repeats. The Kraken scanner is built to run from cron a couple of minutes after each hour closes. The rest is delegation the way you’d delegate to a person: “Figure out why the bot lost money overnight.” “Build the next book in the series.” “What do I need to do to get an EIN and a business license?” I review, decide, and do anything irreversible myself: placing trades, opening accounts, publishing, spending money. What’s worked Narrow agents beat one big chat. A focused agent builds up context in its area. My books agent knows the house style, the folder conventions, and every idea I’ve already rejected it keeps an ideas log . A generalist would have to relearn all of that every time. Make them show their work. My best agents label their confidence. The capital agent tags each claim as verified or not. The trading competitions doc marks unconfirmed figures as UNVERIFIED . The bot-diagnosis report splits P&L into buckets so I can check the math. When an agent tells you how it knows something, you can trust it with more. Alert-only, human-in-the-loop for money. The Kraken scanner can’t trade, and the market-making bots start in dry-run and then go live with a few dollars before I add capital in steps. Losing $1.14 to a bug is tuition. Losing $1,140 is a problem. Real artifacts, not chat answers. Everything worth keeping ends up as a file: a repo, a CSV, a markdown report with a date in the filename. That makes the work reviewable, diffable, and reusable by other agents. What hasn’t Building is easy. Shipping is the bottleneck. Fourteen books are built, and only two are live. The agents can produce faster than I can review, upload, and market. The constraint moved from making things to deciding and finishing. Bots do exactly what you told them to. The ladder bot wasn’t broken in a dramatic way. It followed rules that looked sensible one at a time rotate out of losers, seed new positions with market orders and added up to churn and hidden fees. Agents help you find those bugs faster, but they won’t stop you from writing them. Some things don’t have APIs. Kraken Prop has no trading API, and a lot of bank and bonus offers need a human to sign up, verify, and move money. Some of the “automation” is really just very good checklists. Takeaways you can copy 1. One agent per job. Give each one a name, a scope, and a home folder. Don’t mix your taxes and your side project in one context. 2. Make them write files, not just answers. Dated reports, CSVs, and repos are things you and other agents can build on. 3. Ask for confidence labels. “Mark anything you couldn’t verify” is the most useful sentence I’ve added to my instructions. 4. Ask for “only what’s new.” For anything that runs on a schedule, have it diff against what it already told you. 5. Keep money and irreversible actions human. Use alert-only tools and dry runs, start with tiny capital, and write the pass/fail criteria before you scale up. 6. Watch for the new bottleneck. Once production is cheap, your time goes to reviewing and shipping. Plan for that. What’s next in this series Each future post goes deep on one project, with the real numbers: - The puzzle book factory: how an agent builds print-ready KDP books, and what two live books have actually sold. - Debugging a $4 trading bot: the full post-mortem on the ladder bot, and whether the fixes worked. - Bonus harvesting with an AI scout: which offers were worth it, and which were a trap. - A breakout scanner for a prop account with no API: paper-trade results. - Generative music, 24/7: building a live stream that chat controls. If you’re building something similar, or want to steal any of this, reply and tell me. I read everything. And subscribe if you want to see whether the $3.71 bot ever makes it back to $4.85. — Blake