One human, one AI: running a 94-app company with agents
LSC-Lab is one person and a set of AI agents. Here is what that actually produces, what it still cannot do, and the numbers behind both. Every figure below was read from our own logs on the day this was written.
The setup
I am 49, from Brazil, with degrees in Business Administration and Mathematics. I am not an IT person: I do not write code and I do not read it. I have worked in real estate since 1999, managing every launch of a land developer in Uberlândia; I was president of the regional land developers' association (2021–2025) and I am vice-president of the local real-estate industry association. I started LSC-Lab to answer one question: how much of a company can one person run if the AI does the execution and the human only decides?
Everything below was built that way. I ask in plain Portuguese, the way I would ask a manager. The agents plan, write the code, test it and report back in business terms: what it costs, what it risks, what it earns. I approve or refuse. Money, risk and the business are my calls; the rest is theirs, and every step leaves a record I can audit without knowing how to program.
The answer we built is not a chatbot. It is an operating system made of three parts:
- One brain. A private MCP server with about 60 tools. Every agent, whatever model it runs on, reads and writes the same state: tasks, incidents, decisions, durable memory with sources, and the company's rules.
- Rules as code. Before an agent pushes code, merges, deploys or touches production, it is required to ask the server for permission: 6,348 such checks in the last week, from 36 different agent clients. The CI policy is versioned; the server says ALLOW or DENY in milliseconds, and GitHub branch protection is the hard stop behind it. "Never shorten backup retention without the owner" is a test that fails the build, not a sentence in a prompt.
- A free-first model router. Free tiers come first (Groq, Gemini, GLM-Flash, Mistral's free plan, Cloudflare). Paid providers are a capped fallback (US$0.75 a day), and since mid-October a capped slice of the Claude Max plan's included API credit carries the heaviest jobs. That is why pay-as-you-go AI stays around US$7 a month while the system runs around the clock.
What one evening looks like
On the evening of 26 October (Brasília time) I asked for one thing: a way to make money that does not depend on me selling. What happened next is in the commit log; the times below are UTC, 27 October.
| Time (UTC) | What happened |
|---|---|
| 00:25 | First commit of Quanto Vale: public price-per-m² pages for every neighbourhood with enough audited listings, plus a paid property valuation (Brazilian appraisal standard NBR 14653-2) at R$49. |
| 01:00 | An automated code reviewer approved with reservations. The agent fixed the real ones (non-atomic site swap, service running as root, missing rate limit) and explained in the PR why the others were false positives. |
| 01:25 | I created an API key in our payment account. The agent verified the account through the API, created the webhook, stored the secrets outside the repository and expired its own test sessions. |
| 02:14 | Live: 1,144 pages, 26 cities, live card checkout, the service sandboxed as an unprivileged user. (Today: 1,339 pages, including 96 Florida cities in USD.) |
| 02:20 | A browser robot visited 134 pages like a person would, on desktop and on an iPhone, and filled the order form. It caught a bug no unit test would: visitors from outside Brazil were being charged in US dollars. Fixed by 04:13 and verified against the payment API. |
Under two hours from first commit to a product ready to take money. That is the part that works.
What does not work yet
- Revenue. We cannot yet show recurring revenue in our own systems: one license of another product is active, but its R$49 payment was never recorded as paid. The new product launched with zero sales and is not yet indexed by Google. Building is no longer the bottleneck; distribution is.
- Autonomy. Only 16% of the logged work in the last week was started by robots. The rest started from a conversation with me. "AI works 24/7" is true for execution, not yet for initiative.
- Self-improvement. We run a recursive self-improvement loop: variant prompts and policies compete head to head in an arena every night, and only winners are promoted. So far: 11 promoted, 1,262 rejected. Seven of the winners were policy variants tested on 218 to 1,534 cases; four were skill edits tested on only 5 or 6. The recursive step, a new version of the improver that beats the old one, has not happened: both second-generation improvers produced zero winners (0 of 125 and 0 of 40), on samples too small to tell worse from unlucky. We call the target "level 5". We are not there.
- Mess. 45 incidents are open, 6 of them high severity. Most were found by our own monitors, which is the point, but they are open.
Five things we learned the hard way
- A green check is not an answer. Our CI once reported success on a test suite that had been skipped. Every measurement now has to say what it measured, and every new check is born with a case it must fail.
- The AI repeats mistakes unless memory is a system. The same network bug showed up four times in different places. Durable memory with sources, read at the start of every session, is what stopped it.
- Route by cost, and cap the paid side. No single free tier is reliable, and free models alone did not carry our heaviest jobs. A router with fallbacks across providers and hard daily caps on the paid ones is what keeps the bill small and predictable.
- Guardrails beat heroics. "Production untouched", "cost cap", "no direct push to main" are simple rules. They are why an agent working at 3 a.m. does not take the company down.
- Test like a human. The dollar-pricing bug passed every unit test. It only appeared when a robot used the real site from another country.
What we are doing next
Distribution, measured the same way as everything else: which pages get indexed, which visitors order, what each sale costs. And the self-improvement loop, judged by the same rule: no promotion without samples.
If you are building companies where one person decides and agents execute, or you want to see the system up close, write to us. If you just want to know what your apartment is worth in São Paulo, Rio, Goiânia or Uberlândia, start here.