AgenticRL-Bench (Verifiable)

AgenticRL-Bench, from Eigent AI's Env-Lab, tests whether coding agents can do the engineering behind agentic reinforcement learning: training data and environment selection, harness and inference-engine integration, rewards and verifiers, trainers, monitoring and evaluation. Each task is rebuilt from an incident in a real RL run as a Harbor task with a sandbox image, CPU or one to two H100 GPUs, an 8-hour agent budget and hidden verifiers. This is the Verifiable track's 19 tasks, each given 3 independent attempts; pass@1 is the share of the 57 attempts in which every check passes, and an attempt cut short by an infrastructure failure counts as not solved. Scores are for a model in its vendor's coding harness at a fixed effort. The three open-ended tasks, scored in their own units against a base model, are not included.