AReaL2.0 enables self-evolving LLM agents via online RL at enterprise scale

A new paper from researchers including Ran Yan and Yi Wu argues that the biggest bottleneck for self-evolving LLM agents isn't reinforcement learning algorithms themselves but the lack of proper agentic RL systems at enterprise scale. Current deployed agents—coding assistants, customer-support chatbots, scientific research assistants—remain fundamentally static: weights, prompts, and tools are frozen until a human manually retrains and redeploys them. This prevents agents from improving from their own experience in production.

The authors identify three essential missing pillars: a standardized agent trajectory data protocol that carries RL learning signals at step granularity across heterogeneous agent paradigms; an enterprise-grade data proxy that converts real workloads into governed learning substrates; and a unified agent evolution control plane that automatically decides when to update policy weights or evolve in-context harnesses based on trajectory statistics. They instantiate one branch of this vision as AReaL2.0, which reorganizes existing RL infrastructure into an agent-oriented online training loop that continuously improves agent policies from deployed workloads without human intervention.