Agent entrepreneur
I wanted to see how far an agent can go on its own when you hand it a real, small business. It looks for ideas, builds small tools, publishes them and watches what works. Anything that costs money goes through me.
この事例はフランス語と英語のみで公開しています。
The idea
Most agent demos stop right where things get interesting: when you have to keep going over time, absorb mistakes and report back. So I gave an agent a real setting, with a budget, tools and a logbook.
The loop
Every day, the agent starts from its logbook: what was published, what earns money, what breaks. It proposes what comes next. Decisions that cost something wait for my approval; the rest moves on its own.
What it produced
- tools published on a marketplace
- 7
- automated tests
- 117
- health check every day
- 1
Scraping and automation tools, a template pack for an automation platform, and market monitoring. Nothing spectacular revenue-wise, and that was never the point: I wanted to see where autonomy breaks.
Guardrails
An agent that publishes without tests ends up publishing anything. So every tool has its tests, a health check runs every morning, and a decision gate blocks any spending without my approval.
What I took away
Autonomy is earned in steps. The more precise the logbook, the better the agent’s decisions. The hard part is not getting it to act, it’s knowing when to stop it.