XNews

Hugging Face Holds Training Agents IRL Event at GitHub Offices in San Francisco

Hugging Face Holds Training Agents IRL Event at GitHub Offices in San Francisco

Hugging Face held the in-person closing session of its Training Agents series at GitHub's San Francisco offices, featuring four talks, a panel and audience Q&A spanning nearly two hours on post-training agents with open models. Speakers from Microsoft, Mercor, RadixArk and BenchFlow covered enterprise reinforcement learning, industry environments and benchmarks, training agents with Miles, and the OpenEnv Arena. Full recordings and slides were released.

Event version 1.

Key insights

Enterprise RL priorities

Large enterprises in industrial, manufacturing and energy sectors buy on business metrics such as latency, accuracy and controllable post-training risk. They care about the full chain including tools, data, network, harness and wall-clock time rather than a single model, and real agents often connect dozens to hundreds of tools.

Training versus evaluation environments

Training environments provide rewards, while evaluation environments provide measurement of stability and predictive power for deployment. Agents should be evaluated as model plus harness plus operations, not as bare model scores, and benchmarks should be maintained as long-term products.

Asynchronous agent RL with Miles

Miles treats multi-step environment trajectories as training samples and connects to OpenEnv reset and step interfaces. In a case with a large MoE on Terminal Bench 2 using 64 B300 chips split evenly between inference and training, fully asynchronous with a new sandbox per episode, reward rose from about 0.44 to 0.56 after roughly 100 steps.

OpenEnv Arena environment filtering

OpenEnv Arena screens trainable environments. Of roughly 90,000 findable RL environments,层层 filtering left about 800 relatively clean ones, and about 1% of cheating tasks can contaminate an entire subsequent training round.

Adoption barriers discussed

Stated barriers include a lack of vetted open environments, enterprise model fatigue, result pressure and budgets consumed by coding agents, and atrophied data and MLOps capabilities after the large model API era.

Source posts