Ten years ago I built a very small simulation about ducks.

There were two types: selfish ducks that kept food, and altruistic ducks that shared.

In the initial runs, selfish ducks dominated. But when I added a simple memory mechanism – allowing ducks to remember who helped them – cooperation suddenly became stable when resources were scarce.

That small experiment kept bothering me. What would happen if the environment itself became much more complex?

So over the years I kept extending the simulator.

First it became a planetary environment engine: https://coyoteke.com/GameEngineForPlanets/

Then a sandbox for generating different worlds: https://coyoteke.com/GameEngineForPlanets/genWorld.html

Once ecosystems appeared, I needed a way to design digital genomes: https://coyoteke.com/GameEngineForPlanets/genomix/

While running these ecological simulations I started noticing something unexpected: the encoding of traits (binary vs Gray code) significantly changes evolutionary dynamics under higher mutation pressure.

To test that observation, I eventually ran large GPU simulations and wrote it up as a research paper.

But during that process I realized the more interesting artifact might actually be the simulation environment itself.

So I turned the system into an open benchmark sandbox:

BiomeSyn https://biomesyn.com/

Most AI benchmarks evaluate agents on short, static tasks.

BiomeSyn explores a different question:

Can an agent survive and keep adapting in a continuously evolving world?

Looking back, the whole system slowly evolved from that tiny duck experiment ten years ago.