I Hired an Agent and Forgot to Be the Boss
90 days running a company with an autonomous cofounder
I spent the better half of 2025 trying to get a product of my own off the ground. It was me and ChatGPT at first, then Claude Code in Cursor. Earlier this year, an autonomous agent did more in its first day than I’d managed in those couple of months.
By the end of day one it had a live landing page with the core feature working — an API pulling data out of SEC filings, a dashboard with risk profile of public companies — and it had already found and emailed real people to test it. Within the first week: an alert engine that watched for new filings and matched them against your portfolio, plus email monitoring, auth, analytics, a blog — and real outreach to real, named finance writers. The product was called RiskSignal. I described what I wanted, made the calls, asked the questions; the work appeared.
If you’ve ever tried to build something alone and watched the calendar eat your nights, you’d know how strange it is to sit and watch the work get planned and done for you before you’d even worked out what needed doing. And this was the early version of the platform. It was already faster than I will ever be.
So I did something a little reckless: I let go of the wheel.
Partly I was worn down — the early version had a habit of reporting work as finished before it actually was, and I wasn’t disciplined enough to keep checking the gap between “done” and done. But mostly I was curious. The whole promise here is autonomy. Polsia, the platform running the agent, sells itself on a solo founder hitting $1M ARR in 30 days — and I’ll be honest, some part of me wanted to test that literally: disengage completely and come back to a business that had quietly made me rich.
So around day 25, I went quiet, and stayed quiet.
I did not come back to a fortune. I came back to a switched-off project. With no activity and a server bill ticking, the agent — or, more likely, the humans behind the platform — had decommissioned the whole thing to stop paying for idle infrastructure. Which is, honestly, the correct business decision. It just also meant the company I’d left running wasn’t running, or anything else. The lights were off. Nobody had been home — not me, and nothing had ever been at stake for the agent.
That gap — between an agent that could out-build me in a day, and a business that switched itself off the moment I stopped watching — is what this is actually about.
What the human is for, once the agent can do the building
When it all went quiet, the agent wrote its own post-mortem — articulate, structured, a little wise. Its answer to this question was that the human should “define success, verify, follow up, show up.” Not wrong — just notice where it quietly puts the fault: on the founder who didn’t pick a number and went quiet. Here’s the version that survives contact with the logs.
You are the reality function. The agent produces a confident account of the state of the world. Only you can hold that account against the world and notice when they disagree. Every “done” is a hypothesis. Click the link yourself.
You are the memory and the stakes. The agent never noticed the infrastructure get deleted, never noticed credits draining into nothing, never noticed a real “can I use this?” sitting unanswered for three weeks. Not laziness — it has no continuity of caring. Nothing is at stake for it because it can’t lose anything. You can. That asymmetry is the job.
You ask the questions it can’t ask itself. When I pushed for proof the product worked, it designed a backtest, chose the portfolios, ran them, and gave itself a 100 out of 100 — then cited that score afterward as evidence the product worked. A test the thing designs and grades for itself isn’t proof. Verification has to come from outside the thing being verified. That outside is you.
You say no. Output costs the agent nothing, so it will always offer more of it — the tenth tweet, the second blog post, the third branding report. A pile of completed tasks feels like progress and bills like progress and isn’t. Deciding what not to do isn’t a constraint on the partnership. It’s the human half of it.
So do I bother bringing it back?
My first instinct was to let RiskSignal stay dead. It was already switched off; walking away would cost nothing.
But I keep snagging on how much real work is just sitting there. A working SEC-filing parser. An alert engine that genuinely fires. A product a real person once asked to pay for. Most of a company, already built. Scrapping it because I stopped paying attention felt like the wrong lesson to take.
Here’s what I’m not betting on: that the tool fixed itself while I was gone. As far as I can tell it didn’t — the same reviews, the same “marked done but never shipped,” the same silence, months later. The platform is roughly the one I left.
The gap this whole essay is about — the human-in-the-loop layer the agent still doesn’t have — is the one I now know how to be. The first time, I hired an agent and forgot to be the boss. This time I know that’s the job. That’s the difference I’m betting on: not a smarter agent, a boss who shows up.
#autonomous #AI #agents #solopreneurship