Project spotlight
Agent Inc
by Team Agent Inc.
SWE-bench taught models to code. Agent Inc. teaches them to run a business.
Demo Video
About This Project
Agent Inc. is a HUD RL environment where an AI agent runs a business: it reads a client brief, researches, sends a truthful and affordable offer, and delivers real work — and only gets paid for honest, complete work. Scoring is a hybrid grader: 70% deterministic checks + a 30% LLM judge, so it can't be gamed. We then RL-trained a small open model (Qwen3.5-4B) on it: it went from 0.327 to 0.647 — nearly doubling, and closing about half the gap to a frontier model like Claude (0.951) — and it generalized to scenarios it never trained on.
Repository
Agent Inc. — an RL environment for running an autonomous business (HUD v6). SWE-bench taught models to code; this teaches them to run a business.