AReaL 2.0 makes online RL something an agent can point an endpoint at

LLM agents are moving into production: coding, research, tool use, customer support, automation. But most agents are still static after deployment. AReaL 2.0 is here🔥Introducing RL as Micro-Service for online learning agents, focusing on the missing systems layer: turning real agent workloads into governed, replayable, trainable online RL loops. This is the community's first major move after @AReaL_AI joined the @PyTorch Ecosystem — and a big step toward open, collaborative infrastructure for agentic RL. #opensource #agenticRL #AReaL #ReinforcementLearning

🤔What does "RL as Micro-Service" actually mean? ⬅️Before: agent app ↔️ RL training = heavy engineering bridge ➡️Now: same agent, point to AReaL endpoint, get online learning loop with trajectory capture + async training + hot weight swap With AReaL 2.0, RL components are decomposed into micro-services: Training Service, Inference Serivce, WeightUpdate Service and Agent Service. Existing agents can connect by switching their model endpoint to an AReaL-managed entry point.

AReaL 2.0 is not just a release — it is an open invitation. 🤗If you are building agents, RL systems, inference backends, training stacks, hardware platforms, or evaluation tooling, we’d love to collaborate. Join us in building open infrastructure for self-evolving agents. 🧠 GitHub: https://t.co/HCht8xOCe7 📓 Tech Report: https://t.co/B9OIyTzGRF