What it is
GP-Bench
GP-Bench is the environment we are building for the next test: a solo GP agent runs a $50M pre-seed fund for 40 quarters, in a market that keeps moving.
A static benchmark asks about one founder at one moment. A fund is time, sequence and consequence: deal flow arrives, information costs money, early bets shape later ones, and exits come years after the decision. GP-Bench turns that into a world an agent can act in.
Every episode is generated fresh, so there is nothing to memorize.
How it works
In three parts
- I
The world
Deal flow, the cost of screening, portfolio decisions, follow-ons and exits, in a market that changes as the quarters pass.
- II
The agent
Any decision-making model can take the GP's seat: the same reasoning methods we test on single founders, now running a fund end to end.
- III
The run
Forty quarters, one fund, every decision on the record: which deals were seen, which were funded, what was learned, and how the portfolio came out.
In numbers
- $50M
- pre-seed fund under the agent's management
- 40
- quarters in a run
- Fresh
- every episode, so nothing can be memorized
Benchmarks
Simulation
Harness
Predictions