HN Showcase
Back to browse

A deterministic ecosystem simulator for long-horizon AI agents

by yangkecoy·Mar 6, 2026·1 point·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerBig BrainWizardryZero to One

Deterministic multi-agent evolutionary benchmark with SHA-256 reproducible capsules for agent testing.

Strengths
  • •Deterministic replay via SHA-256 capsules—agents are hashable, shareable, and fully reproducible across seeds
  • •8 frozen scenario suites isolate distinct selection pressures (collapse, scarcity, mutation, entropy)—rigorous benchmark design
  • •Novel research finding: mutation-intensity crossover where Gray encoding loses locality under scaling—publishable insight
Weaknesses
  • •Early-stage platform (v1.0)—limited public agent test data and no comparison to standard benchmarks like ALE or Atari
  • •Niche audience: AI researcher/RL specialist only—not actionable for most agent builders
Category
AI/ML
Target Audience

AI researchers benchmarking adaptive agent behavior, evolutionary algorithm researchers

Similar To

OpenAI Gym / Gymnasium · ALE (Atari Learning Environment) · MLCommons benchmarks

Post Description

Ten years ago I built a small evolutionary toy experiment with two types of agents: selfish and cooperative “ducks”.

At first, selfish strategies dominated. But when agents were given memory — the ability to remember who helped them — cooperation suddenly became stable under resource scarcity.

That experiment stayed in the back of my mind for years.

Recently I started rebuilding the idea from scratch as a larger system:

BiomeSyn

https://biomesyn.com/

Instead of evaluating AI on static tasks, the goal is to explore long-horizon adaptive environments where agents must:

• gather resources • survive environmental pressure • compete with other agents • adapt over many generations

The system is deterministic, so experiments can be reproduced across seeds — which makes it possible to treat it as a benchmark for adaptive agents.

The bigger question I’m interested in:

> What happens when intelligence is evaluated inside a world that keeps evolving?

Many current benchmarks measure short-episode performance. But real adaptive systems must operate in open-ended environments.

BiomeSyn is still an early research sandbox, but I’m curious whether environments like this could become useful for studying:

• evolutionary computation • long-horizon RL agents • multi-agent ecosystems • adaptive AI systems

Would be interested to hear thoughts from people working on agents, simulation platforms, or large-scale AI systems.

https://biomesyn.com/

Similar Projects