AGI Maze is a framework for evaluating AGI agents and their ability to build and use world models in partially observable environments.
In this context, “world modeling” is meant broadly:
- not just discovering simple regularities in an environment,
- but forming usable descriptions of partially observed worlds and reasoning about them.
This repository contains a clean, public-facing package of materials for developers and researchers:
- a stable, developer-facing API specification
- baseline agent examples (reference implementations)
- benchmark utilities (small runners + reporting scripts)
- API — the HTTP+JSON interface for starting episodes, taking actions, and receiving observations.
- Agents — an overview of included baseline agents and shared helpers.
- Benchmark tools — lightweight utilities for repeatable runs and quick statistics.
- Results — calibration notes and baseline results (where applicable).
docs/— documentation (API, agents, benchmark tools, results)agents/— baseline agents and example clientsbench/— benchmark runner/report scripts and helpers