Show HN: FLE v0.3 – Claude Code Plays Factorio (jackhopkins.github.io)

The Factorio Learning Environment (FLE) v0.3.0 release wires frontier agents — notably Claude Code — into Factorio for large-scale, long-horizon planning research. Key upgrades: FLE no longer needs the Factorio game client (a headless renderer now supplies realistic pixel observations for multimodal models), the API conforms to the OpenAI Gym interface for easy integration, and a CLI plus cluster orchestration enables 1‑line experiment runs and massively scalable evaluation. The project is open-sourcing evaluation tooling (Weights & Biases logging, sweep resuming, analysis) and livestreaming agent runs to showcase capabilities in interactive, multi-turn tasks. Technically, agents operate via a programmatic action API (move_to, place_entity, connect_entities, set_entity_recipe, insert_item, extract_item, etc.) and receive rich Observation objects containing raw_text, entities, inventory, research state, game_info, production flows and inter-agent messages. The release demonstrates a Claude-driven agent constructing an iron gear wheel factory end-to-end—discovering water for power, sizing mining throughput (drill/furnace math), placing assemblers, wiring power and belts, and iteratively debugging blocked inputs—illustrating long-term goal setting, world modeling and dynamic recovery. For AI/ML researchers this means an accessible, reproducible benchmark for multimodal, multi-agent, and long-horizon planning research with realistic, programmatic environment interaction.