Specialization becomes agents' RL playground

Labelbox launches Recursion, a platform using reinforcement learning to train specialized AI agents that outperform larger generalist models.

With Recursion, Labelbox aims to transform enterprise workflows into training environments for specialized agents.

The platform unifies development, evaluation, and deployment around a reinforcement learning loop, powered by proprietary data, internal tools, business rules, expert feedback, and edge cases from each organization. Labelbox emphasizes a simple principle: for specific tasks, a smaller model trained on the right domain can outperform a larger, generalist model.

The company notably cites a 35B model fine-tuned with Recursion, outperforming a 397B model on agentic financial tasks held out from training, as well as a support agent demonstrating a better resolution rate, fewer hallucinations, faster responses, and a lower cost per ticket. The WorldSim component serves to recreate realistic software environments, while evaluations measure final outcomes, intermediate decisions, and execution quality.

Recursion thus positions itself less as an out-of-the-box agent tool than as a machine for converting internal expertise into specialized models that improve with use.