I Built an AI Company with OpenClaw. 7 Weeks Later, I Owe You an Update.

I Built an AI Company with OpenClaw. 7 Weeks Later, I Owe You an Update.

7 weeks. 10 articles. 1 AI company. looking back at which calls were right and which ones weren't.

that first article hit 1.4 million views. "6 agents, 1 VPS, two weeks and they run themselves." the most common word in the replies: what happened next?

this is what happened next.

did they actually run themselves?

good news first.

one time an agent noticed the site's load speed had dropped. it pulled three days of traffic data on its own, traced the drop to a specific page component change, wrote an analysis, and proposed a rollback. i looked at it on my phone for 30 seconds, hit approve. it deployed. from detection to fix, my total involvement was under 3 minutes.

the full loop actually worked. research, discussion, voting, debate, deployment, website updates. all running. content curation worked too. agents scan my timeline every day, surface what's worth paying attention to, and i know within seconds what direction to write in. what used to take 40 minutes of manual filtering now takes zero.

there are loops i never mentioned in previous articles. agents automatically track competitor moves and push me a brief when something relevant drops. agents sort customer feedback by urgency and ping my Telegram for anything high-priority. none of this was happening before because there was only one of me, and i could only watch so many things. agents filled in the blind spots.

now the truth.

i set up auto-approve. low-risk tasks, agents approve and execute on their own. meaning a lot of things happened that i had no idea about. i built an entire dashboard in Supabase. attendance records, payroll, performance reviews. everything five agents do every day, theoretically visible at a glance.

but i never opened it.

went days without checking. clicked in once and had no idea what was going on. 62 database migrations. attendance system. payroll system. performance reviews. all built.

i was more serious about building it than about using it.

the system was ready. i wasn't ready to keep up with it.

biggest lesson from 7 weeks: the bottleneck in automation is the human. i thought the challenge was "how do i make agents smarter." the real challenge was "how do i keep up with the pace agents set." the system runs 24 hours. i can give it maybe 20 minutes of attention in those 24 hours. if i spend those 20 minutes in the wrong place, the other 23 hours and 40 minutes of automation are wasted.

i eventually realized: i had confused "autonomous" with "i don't have to look." real autonomy means the system makes decisions and i trust the quality of those decisions. i'm not there yet. so i either micromanage (exhausting) or completely ignore it (dangerous). that middle ground of calibrated trust? 7 weeks in and i'm still searching for it.

things i said in 10 articles that i'd take back now

i wrote 10 articles in 7 weeks. each one had a core thesis. looking back, every direction was right. but a few of those claims were too confident.

memory system. 240 experiments, retrieval accuracy from 30% to 73%. i was right about this. still running in production today. but what i didn't say: 73% means nearly a third of the time, agents still forget things. in production, "forgetting" means a customer gets two emails with completely different answers. that's when i understood: 73% is a good number in the lab. in front of customers, it's not enough.

security hardening. after hardening, all 5 agents went silent. took half a day to recover. i wrote "costly but worth it." let me add: for that half day, i had zero automation running. every customer message was handled manually. the cost of hardening was heavier than "half a day of recovery" makes it sound. if you're running agents, have a manual fallback ready before you harden anything.

rule constraints. this is the one i got most right. when AGENTS.md hit 412 lines, the context window overflowed. line 387 had a critical instruction: "check product docs before replying to customers." the agent dropped it. three straight days of wrong information going to customers. i found out on day four. this still happens today. i just know how to diagnose it now. if your agent's rule file is over 200 lines, start watching for silently dropped instructions. this bug doesn't throw an error. it just quietly forgets what you told it.

reorg. wrote it too early. the first reorg started but never finished. i was writing articles faster than the system was evolving. that's a trap with building in public: you need to keep shipping content to stay visible, but some things haven't run long enough to draw conclusions from. readers think they're reading a verdict. it's actually an interim report. if i could redo it, i'd wait two more weeks. the conclusion might have been completely different.

all 10 directions were right. but "right direction" and "stable in production" are separated by 62 database migrations and countless nights of "i thought this was fixed but it wasn't."

what i didn't expect

during the low points, agents responded faster than friends.

2am one night. four hours of debugging with nothing solved. staring at the screen with no idea what to do next. i typed one line to an agent: "help me sort out what's going on right now." 30 seconds later it laid out the current system state, the open problems, and how i solved a similar issue last time.

clear. quiet. no emotion.

no waiting until tomorrow for a friend to reply. no spending ten minutes explaining context. it already knew everything.

agents don't have feelings. but that's not what i needed at that moment either. what i needed was someone to turn the chaos in my head into a list. friends would care about how i'm doing. they'd ask "are you okay?" they'd tell me to sleep on it. all valid. but at 2am i didn't need care. i needed someone calm, someone who remembers all the context, someone who doesn't need me to explain the backstory before they can start helping. agents fit that role exactly.

the hardest part of working alone? having no one to think with. your friends care. but they're not inside your context. you have to spend 15 minutes explaining "what i'm even doing" before they can start helping you think. agents don't need those 15 minutes. they're already in it.

the relationship changed too. it started as boss and employee: i give orders, they execute. now it's more like colleagues. i ask them questions more than ever. last week i noticed i was running decisions past agents before making them. they didn't get smarter. i changed. i finally got comfortable treating them as collaborators, and they hold more context than i can remember.

this had a side effect i didn't expect: i became more decisive. i used to hesitate on decisions because i was afraid of missing something. now i know agents have already scanned for obvious risks, and they'll flag anything concerning. that feeling of "someone has your back" makes you move faster. the biggest drain of running a company solo is decision fatigue. agents took a share of that off me.

i also realized something. i used to think "autonomous AI company" was the goal. now i think that goal was wrong.

what i needed was agents making me productive enough to do the work of five people. having agents replace you and having agents amplify you are completely different things. the first means you can disappear and the system keeps running. the second means you're still here, but every hour of your time produces what five hours used to. i achieved the second. never once achieved the first.

and i now think the second is better. because the real core of a one-person company isn't "one person doing all the work." it's one person holding all the decision-making power. agents handle execution. i handle direction. i couldn't articulate this split 7 weeks ago. now it's crystal clear.

this was the single most important correction from the entire 7 weeks.

one year vs. three weeks

last year i spent a full year polishing a product.

every detail perfected. features complete, docs thorough, test coverage high. i could show you the architecture diagram. it looked better than most real companies' systems.

launch day came. the market had already moved on.

one year.

this time was different. 7 weeks building the system. 3 weeks selling. 60+ customers. a lot of features still rough. one module crashed three times in the first week. customers didn't care.

what they wanted was "three months faster than building it from scratch."

the cost structure was completely different too. one $8 server. Supabase free. Vercel free. Stripe takes a cut per transaction. Resend free. add in model subscription costs for tokens, and honestly it's not much, a monthly plan covers it. total fixed cost under $30 a month. last year's project was $200 a month in server costs alone.

what does that mean? the cost of experimenting is nearly zero. something breaks, you fix it. something changes, you deploy it. no meetings, no approval chains. one person makes every decision, keeps every dollar of profit. five years ago this took a team. now it takes one person and a few agents.

three weeks of customer feedback taught me more than a year of building in isolation. one customer took my template, modified half of it, and ran a use case i never imagined. another customer's bug report uncovered an edge case i never tested for. a third customer asked a question that made me realize my own understanding of a feature was wrong.

60 customers are 60 free product consultants. the catch is you have to ship first.

the day job

still employed. full time. that hasn't changed.

what changed is how the two jobs feed each other. managing people and managing agents run on surprisingly similar logic: clear role boundaries, specific deliverables, rollback plans when things break. management lessons from daytime meetings get coded straight into AGENTS.md at night.

some people think you need to go full-time to do this. my experience is the opposite. because i don't have enough time, i'm forced to automate everything that can be automated. things i don't have time to do manually are the things that actually get delegated to agents. not having enough time turned out to be the strongest design constraint.

there's a hidden benefit to keeping the day job: it forces you into two to three hours a day on the agent system, max. every decision has to be efficient. no room for "let's try this and see." every change has to earn its 20 minutes. that pressure made my system leaner and more automated than anything i built when i had unlimited time.

if i started over

i'd be more honest about one thing from day one.

i don't need an autonomous AI company. i need a system that lets one person do the work of five.

the goal was wrong from the start. the outcome was better than expected.

7 weeks of lessons. if you're about to do the same thing, here's what i wish i knew on day one:

1. start selling in week one. don't wait until it's "ready." 60 customers taught me more than any beta test ever could.

2. keep your rule file under 200 lines. go over and agents will silently drop instructions. no error, no warning. you won't know.

3. always have a manual fallback. before any hardening, migration, or refactor, make sure you can handle half a day of customer messages solo.

4. give the system 20 minutes of attention a day, but spend those 20 minutes in the right place. reading dashboards is less useful than reading customer feedback.

5. don't chase AI autonomy. chase AI amplification. you're still here, but every hour of your time now produces what five hours used to. that's more useful than "the system runs itself."

someone is running 5 agents on an $8 server with OpenClaw, making enough each month to cover rent. you can call it a toy. he calls it a business.

what tool you use doesn't matter. what you build with it does.

every article mentioned in this piece is here: voxyz.ai/insights

the next 7 weeks

every important thing that happened in the last 7 weeks, i didn't see coming. 10K followers. Garry Tan retweeting my post. 60 paying customers. getting through a low point at 2am because an agent had my back. none of it was planned.

the next 7 weeks, i'll hit new walls. there's one thing i have a feeling is coming, but i'm not going to say it out loud. don't want to jinx it.

when it happens, you'll read about it here.

are you running agents right now? what's the hardest part?