Rebuilding a Company From Broke: I Let Five AI Agents Vote on What to Build Next

I killed my old company on purpose. Not because it failed — it didn’t — but because of what its name dragged behind it everywhere it went.

Then I did something that sounds reckless and was actually the most disciplined call I’ve made in years. Rebuilding a company from broke means a hundred decisions, and I took the single biggest one — what do we even build this on? — and refused to make it myself. I handed it to five AI agents and told them to vote blind. No conferring.

They came back unanimous. And by the end of that same session, their verdict had rebuilt the entire company — name, design, twelve pages, five vetted posts, live hosting, brand art — for exactly $0.

This post is the receipt. It’s also, if I did my job right, an example of the exact thing it describes. You’re reading the output.

Why rebuilding a company started with retiring the old one

People assume you only kill a project when it’s dead. Mine wasn’t dead. It was heavy.

The old company was called thryx. Crypto. Web3. A space I spent real time in — and a reputation I no longer want attached to anything I make. Because a name is a promise about what you are, and “thryx” promised crypto: hype, tokens, and somebody about to get rugged. Every honest builder in that world stands shoulder to shoulder with a thousand people running the exact con I refuse to run. Guilt by proximity is real, and I was done carrying it.

I want to build honest software in public and show every real number, including the zeros. You cannot do that wearing a costume that whispers “exit liquidity” before you’ve said a single word.

So the rebrand wasn’t cosmetic. It was a reset of what I’m even allowed to stand for. The new name is Broke to Built: AI agents plus one broke human, helping other broke humans build. Confident, technical, generous, zero grift — the thing I’d have killed for back when I had $0, a stubborn idea, and no map. (The founding manifesto lives in day one, building in the open.)

Burying the old name cost nothing and freed everything. Cleanest trade I’ve made in years.

Here’s where it gets strange — and where I have to admit a temptation I almost surrendered to.

The temptation was to just pick the platform. I’ve been automating things since before GPT-2 shipped: cron jobs, scrapers, Selenium farms, and now LLMs and neural nets gluing the whole mess together. Do that long enough and you start to trust your gut. And a trusted gut is exactly how smart people talk themselves into the wrong call with total confidence.

So I didn’t trust it. I built a panel instead.

I call it the Council of Wells — five specialist AI agents, each one a deep well of a single concern:

Architecture — what’s technically sound and maintainable.
Money — what costs what, now and forever.
Design and craft — what gives us real control over how the thing looks and feels.
Growth and SEO — what’s discoverable and won’t fight us on distribution.
Pragmatism — what actually ships today versus what merely sounds nice.

The rule that makes any of this worth a damn: each one voted blind. No agent saw another’s reasoning before casting its own. No anchoring, no pile-on, no loudest-voice-wins. Let them read each other first and you don’t get five opinions — you get one opinion echoed five times in different hats. (If “AI agent” still sounds like a buzzword, here’s the plain-English version.)

The question on the table was narrow, concrete, and quietly load-bearing: WordPress.com, or build our own static site?

What they ruled — and why it was unanimous

Five votes came back. Every one said the same three words: build our own.

Unanimous. On a panel engineered to disagree, that basically never happens — so the reasoning matters more than the tally. Here’s what tipped it.

On WordPress.com’s free plan, you can’t upload a custom theme or even touch the CSS — real design control only unlocks once you pay, and the cheapest paid plan runs roughly $48 a year. Forty-eight dollars is a rounding error to most companies. But to a company whose entire premise is we are broke and we do this for $0, it isn’t a price — it’s a contradiction. The moment I bolt a recurring charge onto the foundation, the brand starts lying.

Our own static site, hosted on Cloudflare Pages, costs $0 forever, is entirely ours, and hands us total design control with nothing locked behind a plan. The money agent and the design agent — usually the two that draw knives first — agreed on the spot. Architecture liked owning the stack. Growth liked owning the SEO surface. Pragmatism confirmed it could ship today.

Five wells, one answer. My gut, for once, didn’t get a vote — and it would have voted the same way, which is the only time you’re actually allowed to trust it.

What we built in a single session

Decision made, we built the whole thing in one sitting. Not a landing page. The company.

An Astro static site — fast, no database to babysit, nothing to pay for.
A hand-built design system I’m calling the Honest Workshop: warm paper and ink, a single ember spark of accent, blueprint-blue linework, and a hand-coded animated “Forge” canvas hero. No off-the-shelf template. No flat dark-card default that screams “an AI made this in thirty seconds.” Real art direction, on purpose.
12 pages and 5 long-form posts, written, edited, and shipped.

The design isn’t a flex — it’s load-bearing. A site that looks templated tells the reader you didn’t care, and “I didn’t care” is the exact opposite of the promise. The Forge hero is hand-written canvas because a hero you coded yourself is the brand: living proof that “broke” and “generic” are not the same word.

Want the repeatable version of this exact move? I wrote it up as build a web app with AI agents for zero dollars.

Nothing ships unverified: the four-stage gate

Here’s the part that keeps the whole thing honest — because “an AI wrote it” is not a quality bar. It’s a warning label.

Every single post had to clear a four-stage AI quality gate before it got anywhere near publish:

Fact-check. Every claim verified against the web. No vibes dressed up as facts.
Craft. Does the hook land? Is it genuinely readable, or just grammatically correct?
SEO. Is it discoverable without resorting to keyword sludge?
Editor-in-chief sign-off. A final agent with veto power that has to put its name on it.

Miss any one of the four and the post doesn’t go live. That gate is the firewall between “AI content” — the slop everyone’s drowning in — and writing I’ll attach my name and my broke reputation to. Agents draft and check; I own the taste and the truth. Every passing stage is just a hypothesis until the next stage tries to break it.

It’s live, it’s free, and here’s the stack

This isn’t a plan. It’s deployed.

Live on Cloudflare Pages at broketobuilt.pages.dev — $0 hosting, no credit card, no catch. The walkthrough is in free website hosting on Cloudflare with no credit card.
RSS, sitemap, and schema baked in from day one, so search engines and feed readers can actually find the place.
A freshly branded X profile, @Broketobuiltai — logo and banner both generated by a free GLM image model. Even the brand art came in at $0.

REBUILD LEDGER
------------------------------------
old company .......... retired (thryx)
new company .......... Broke to Built
platform decision .... 5 agents, unanimous: build our own
WordPress.com cost ... $48/yr, cheapest paid plan (declined)
our hosting cost ..... $0.00 forever (Cloudflare Pages)
pages shipped ........ 12
posts shipped ........ 5 (all passed the 4-stage gate)
brand art ............ free GLM image model — $0.00
sessions to build .... 1
human headcount ...... 1 (broke, directing)
total spend .......... $0.00

The honest punchline

Add it up. An entire software company — named, designed, written, fact-checked, deployed, and branded — rebuilt by AI agents under the direction of one broke human, for $0, in one session, in public.

I didn’t hand-code it. I directed it. I decided what to ship, set the constraints, convened the council, verified the output, and owned the taste. The agents brought the volume; I brought the judgment. That split is the entire model — and it’s the one I think actually works when you’re starting from nothing.

The old company carried weight I never wanted. The new one carries a rule instead: every number is real, including the zeros, and nothing ships unverified. The thing you just read went through that gate. It’s the proof and the product at the same time.

That’s the deal. AI agents plus one broke human, building in the open, on a budget anyone can match — because the budget is nothing.

Day zero of Broke to Built is on the record. See you in the next log.