Factorio as an LLM torture test
An open-ended benchmark where LLMs write Python to survive Factorio’s industrial complexity.

What it does
FLE is an evaluation framework that drops LLM agents into Factorio, the notoriously intricate factory-building sim. Agents interact by writing Python programs that the environment executes in a REPL loop; they observe the world through the stdout and stderr of their own code, then generate the next program to react. The authors pitch it as a non-saturating benchmark—one that should stay challenging as models improve.
The interesting bit
Most game benchmarks eventually top out. Factorio’s combinatorial resource-management complexity makes that hard. The REPL interface is also unusually raw: the agent observes the world only through the output streams of its previous program, then must synthesize new Python to respond.
Key highlights
- Open-ended evaluation designed not to saturate as frontier models improve
- Agents interact via code synthesis through a REPL: they write Python, the environment executes it, and they read the resulting
stdout/stderr - Built on the actual Factorio game (
v2.0.73+), not a simplified clone - Ships with a CLI, leaderboard, and an arXiv paper for reproducible research
- Optional extras for MCP protocol and PostgreSQL logging
Caveats
- Requires owning Factorio for rendering, plus Docker and Python 3.10+
- README is light on concrete task definitions or API surface details
Verdict
Worth a look if you are building LLM benchmarks or researching open-ended agent evaluation. If you just want a Factorio bot that works out of the box, this is a research framework, not a finished player.
Frequently asked
- What is JackHopkins/factorio-learning-environment?
- An open-ended benchmark where LLMs write Python to survive Factorio’s industrial complexity.
- Is factorio-learning-environment open source?
- Yes — JackHopkins/factorio-learning-environment is an open-source project tracked on heatdrop.
- What language is factorio-learning-environment written in?
- JackHopkins/factorio-learning-environment is primarily written in Python.
- How popular is factorio-learning-environment?
- JackHopkins/factorio-learning-environment has 1k stars on GitHub.
- Where can I find factorio-learning-environment?
- JackHopkins/factorio-learning-environment is on GitHub at https://github.com/JackHopkins/factorio-learning-environment.