Agilerl

AgileRL is a reinforcement learning (RL) infrastructure company that develops an open-source Python framework and a commercial end-to-end platform called Arena. Arena takes a customer's own data, workflows, or environments and converts them into LLM datasets and RL environments, validates them before training, configures algorithms, rewards, constraints, and objectives, runs distributed training across the customer's compute, and supports one-click promotion of checkpoints to production hosted on the customer's own infrastructure, with continual learning from live feedback. The company positions this as an alternative to relying on a general-purpose frontier model, arguing that smaller models fine-tuned on a specific task can outperform generalist models while keeping model weights and training inside the customer's environment.

The underlying open-source library, distributed as the pip package `agilerl` and hosted on GitHub, is described as a deep RL library focused on "RLOps" (MLOps for reinforcement learning). It implements evolutionary hyperparameter optimization — training a population of agents, selecting the strongest via tournament selection, mutating them, and discarding weak performers — which the company says removes the need for many sequential tuning runs and yields roughly 10x faster hyperparameter optimization than existing approaches. The library covers on-policy, off-policy, offline, multi-agent (PettingZoo-style parallel API), contextual multi-armed bandit, and LLM fine-tuning algorithms, including DQN, Rainbow DQN, DDPG, TD3, PPO, CQL, ILQL, IPPO, MADDPG, MATD3, NeuralUCB, NeuralTS, GRPO, GSPO, CISPO, DPO, SFT, LLM PPO, and LLM REINFORCE, plus evolvable network modules (MLP, CNN, LSTM, ResNet, SimBa, GPT, BERT) and distributed multi-GPU training.

The company reports customer and user traction in research, defense, finance, robotics, and logistics, and publishes case studies with RTDynamics (aerial interception missions), Warburg AI (financial RL workflows and compute cost reduction), and Decision Lab (bin-packing/logistics utilization). Press coverage of its seed round states the framework has been used by engineers at Airbus, IBM, JPMorgan, Wayve, and Huawei.

Founding story

Cofounders Param Kumar (CEO) and Nicholas Ustaran-Anderegg (CTO) started the company through the Entrepreneur First programme in the UK ecosystem. Kumar has said the idea came from having built a reinforcement learning system from scratch at a previous company, where he found the process costly and complex, with every new use case requiring simulators, reward designs, and deployment pipelines to be rebuilt. Kumar has also described the post-ChatGPT shift of corporate budgets away from RL toward transformers, and the company's thesis that RL must be layered on top of large statistical models because of the limits of what can be inferred from data alone.

Business model

AgileRL operates an open-source-core model: a freely available Python framework (`agilerl`) that drives adoption, with the commercial Arena platform layered on top for managed, end-to-end enterprise RL training and deployment. Deployment is oriented toward running inside the customer's own infrastructure and compute stack, with model weights retained by the customer.

Tiered software subscription. Business Insider reports a free tier with a limited amount of training credits, paid tiers for professionals and businesses, and custom licenses for larger enterprises.

Traction

The company reports 400,000+ community downloads of its open-source framework on its website, while January 2026 press coverage cited more than 300,000 downloads. The GitHub repository shows 945 stars, 77 forks, and 2,105 commits. AgileRL states its framework is used by leading research labs and institutions and by teams in research, defense, finance, robotics, and logistics; press reports name Airbus, IBM, JPMorgan, Wayve, and Huawei as organizations whose engineers have used the technology. Three customer case studies are published (RTDynamics, Warburg AI, Decision Lab). The company raised $7.5M in seed funding announced in January 2026.

Latest developments

In January 2026 AgileRL announced its $7.5 million seed round and US expansion plans. Recent product and engineering updates highlighted on the site include the Arena Client for running RL at scale from the terminal, an async-RL system the company says beats TRL and ART by 7x, and guidance on selecting RL algorithms and reward systems for multi-turn LLM training. Documentation lists Arena, Arena Client, LLM fine-tuning with GRPO/CISPO/GSPO/DPO/SFT, and multi-frequency population-based training among current capabilities.