Reinforcement‑learning agents need a gym that bundles a task, an executable environment, and a verifier that can reliably tell success from failure. Traditional gym construction is manual, costly, static, and quickly saturates as models improve, leading to contamination. Synthetic generation can scale but single‑pass synthesis yields tasks whose difficulty is mostly superficial; models with comparable capability solve them by exploiting phrasing, and correctness must be judged afterwards by unreliable LLM judges.\ \ AutoGym generates complete gyms—tasks, environments, and verifiers—from a minimal domain seed or prior trajectories. It introduces three mechanisms:\
- Blueprint‑first generation defines the valid solution space, environment requirements, and verification criteria before materializing the environment, making solvability a construction prerequisite rather than a post‑hoc property.\
- Explicit generation parameters steer task topology, interaction depth, capability axes, question obfuscation, and distractor composition, enabling fine‑grained difficulty control.\
- Active curriculum synthesis uses performance‑informed calibration to adjust the distribution over these parameters as model capabilities evolve.\ \ Across productivity and temporal‑reasoning settings, AutoGym produces gyms that span the entire capability spectrum, including instances that challenge state‑of‑the‑art models.\ \ Review