Skip to content

Alien Biology

A framework for testing agentic AI reasoning through procedurally generated biological systems untainted by training data. Alien Biology provides a way to measure complex, agentic reasoning/learning that is:

  • REAL-WORLD - measures performance on practical, complex, real-world-relevant agentic reasoning/learning tasks
  • UNTAINTED - avoids confounding connections to LLM training corpora by drawing tests from an "Alien" universe
  • CONTROLLABLE - is parametrically constructed in ways that allow fine-grained analysis of the limits of agentic reasoning

Documentation

  • Demos - Interactive demonstrations with notebooks, scripts, and output
  • User Guide - Core specs, generators, execution, and agent interface
  • Architecture - System architecture, data model, and protocols
  • API Reference - Auto-generated Python API docs

Quick Start

# Clone the repository
git clone https://github.com/oblinger/alienbio.git
cd alienbio

# Install with uv
uv sync

# Run tests
just test

# Run a scenario
bio run catalog/jobs/hardcoded_test