AgileRL

AgileRL builds tools for training specialized AI agents with reinforcement learning. Its main product, Arena, is an enterprise platform that turns a company's raw data and workflows into training environments and LLM datasets, then trains, tunes and deploys a small model that becomes an expert at that specific task instead of relying on a general frontier model. Arena is aimed at teams building agents in research, defense, finance, robotics and logistics. Users connect and validate their data before spending on compute, select algorithms, rewards, constraints and objectives, and can turn on evolutionary hyperparameter optimization to explore configurations automatically. Training runs on an asynchronous RL engine distributed across the customer's own compute, with metrics, sample efficiency and checkpoints monitored in real time. Reinforcement fine tuning for LLM agents is supported, finished models can be promoted to production with one click on the customer's own infrastructure, and live feedback feeds continual learning. Model weights remain with the customer, and any open source model can be used. An Arena Client brings these workflows to the terminal through the agilerl Python package. Arena pricing is not published; access is arranged by booking a demo, in which the team trains and deploys an agent on a prospect's task in a live session. The underlying AgileRL framework is open source under the Apache 2.0 license and free to use, covering single and multi agent training across on policy, off policy, offline RL, bandit and LLM methods. The company was founded in 2022, has offices in the UK and US, and states that it has raised $7.5 million.