bad.ventures

In build

Most ventures are bad.

We're creating an autonomous venture lab: software for testing business ideas on real customers, and killing most of them cheaply. AI agents are meant to propose; fixed rules and one person decide what reaches the world.

Built today: the rules and approval engine, tested end to end with a scripted stand-in playing the AI. Next: a real AI model, then a first real customer.

The sieve Many ideas enter on the left. A screen refuses most of them, a first experiment fails most of the rest, real money kills most of what is left, and one comes out as a business. screen first experiment real money refused failed killed a business The sieve Many ideas enter at the top. A screen refuses most of them, a first experiment fails most of the rest, real money kills most of what is left, and one comes out as a business. screen refused first experiment failed real money killed a business
How the lab is meant to work. It hasn't met a real customer yet.

01

One door to the world.

How the lab is built The design. On the left, three AI agents in a sandbox: the scout, the venture agent and the reviewer. They send proposals to a wall, the control plane, which holds every rule, cap and record, and refuses a proposal with a reason. The wall has one door, the Gateway. Only what the operator approves goes through it to the world. Evidence comes back from the world through collectors. Today a scripted stand-in plays the agents; no real AI model has run a step on record. AI agents sandboxed · no direct network scout venture agent reviewer today: a scripted stand-in proposes refuses, with a reason control plane every rule every cap every record gateway operator holds the key approves only what is approved the world evidence, via collectors How the lab is built The design. At the top, three AI agents in a sandbox: the scout, the venture agent and the reviewer. They send proposals to a wall, the control plane, which holds every rule, cap and record, and refuses a proposal with a reason. The wall has one door, the Gateway. Only what the operator approves goes through it to the world. Evidence comes back from the world through collectors. Today a scripted stand-in plays the agents; no real AI model has run a step on record. AI agents sandboxed · no direct network scout today: scripted venture agent reviewer proposes refuses, with a reason control plane every rule · every cap · every record gateway only what is approved operator approves the world evidence comes back via collectors
The design. Today a scripted stand-in plays every agent in the tests. No real AI model has run a step on record yet.
  • No agent moves money.
  • Nothing fake.
  • New authority is the operator's.

02

Only behaviour counts.

Research informs a thesis. Only a stranger's behaviour confirms one.

  • verifieda collector
  • attestedthe operator
  • claimedthe agent
The evidence ladder, E0 to E5 From strongest to weakest. E5, money: a settled payment tied to a ledger entry. E4, a stranger pays a cost: a click, a contact, their time or an accepted price. E3, observed by a collector and counted against a denominator. A line marks where the agent's own records stop. Below it, E0 to E2: research and the agent's own claims, which inform a thesis and are never a verdict. E5 Money a settled payment, tied to a ledger entry E4 A stranger pays a cost a click, a contact, time, a price accepted E3 Observed by a collector counted against a denominator The agent's own records stop here E0–E2 Research and claims what the agent read and says; informs a thesis, never a verdict

A $10k-a-year business counts.
So does a big one.
So does a clean kill.

Get in touch.

info@bad.ventures