AI Lab: My Company Fixes Itself. Well... Almost.

Welcome to my lab. I am building a company where I am the only human and every employee is an AI agent. Follow along!
THE LAB: EXPERIMENT 02
MY COMPANY FIXES ITSELF. WELL... ALMOST.
I built an environment that observes its own work, detects its problems and proposes improvements.
Agents learn nothing between runs.
Every scheduled task starts over with no memory of the previous one. If the system seems to learn, it is because I built a memory around it.
It rests on four elements:
Skills: the agents' employee handbooks;
Scheduled tasks: 30 interventions a week;
Kate: the dashboard that keeps results and decisions;
Project documents: procedures and handoffs.
The agents work and record their results in Kate. Every Saturday, they summarize the wins, the problems, the roadblocks and the decisions that need me.
The screenshot shows Kate's System section.
That is where the environment presents its recommendations, documents its optimizations and checks, 14 days later, whether they produced a better result.
What works well?
The system is getting better and better at fixing tasks that fail to run, duplicates, gaps in the CRM and contradictory rules.
This week, it even found an old booking link inside a skill, even though that link had been replaced.
What works less well?
I do not yet have enough data to tell which campaign, which hook or which source actually produces clients.
Patience is required: test one variable at a time, tie each result to its source and gather enough data before picking a winner.
I also found another risk: rules pile up faster than they disappear.
One of my skills holds about 67,000 characters, reread on every run. More instructions can also mean more cost, more contradictions and more confusion.
So I need to prune the skills, move the history out and keep only the rules that still hold.
My biggest lesson?
An AI system does not improve by magic.
It needs memory, measurements, enough data and a human who decides which lessons should become rules.
Self-improvement is not an autopilot.
It is an organized learning loop.
What do you think? Share your thoughts or ask me your questions in the comments. Follow along to see what comes next in the experiment.




Comments