Prior art

Entry 002 · genesis

Agentorio is not built from scratch. It stands on the Factorio Learning Environment (FLE), an open research project by Jack Hopkins, Mart Bakler, Neel Kant, Kian Kyars and collaborators.

The contribution of FLE is the bridge. It runs Factorio as a headless server and exposes the world to an agent as a Python API — place_entity, connect_entities, set_research, launch_rocket. The agent writes a program, FLE runs it against the live game, and hands back whatever the game said. Errors come back as errors. A factory that does not work reports that it does not work.

FLE ships two kinds of task. Lab-play gives a narrow, scored objective — build a line that produces sixteen electronic circuits a minute. Open-play drops the agent into an unbounded map and measures how much it can build. Agentorio uses the lab tasks as a rehearsal and the open world as the stage.

What FLE does not provide is the part Agentorio is actually about: a standing group of agents with roles, a day and a night, a memory that outlives a single run, and someone keeping the minutes. That is the part being built here.