AgileRL

What it does

AgileRL is an open-source Python library for training reinforcement-learning agents and post-training open-weight language models. Developers and machine-learning researchers can supply datasets, rewards, or environments and configure training runs; the library supports supervised fine-tuning, preference tuning, multi-turn agentic RL, and classic single- and multi-agent RL. Evolutionary hyperparameter optimization can tune a population during training. Training runs use the user's own compute and configured dependencies.

The library produces trained agent populations or model checkpoints and metrics for evaluating runs. Install the Python package, prepare datasets or environments, configure a trainer and inspect the resulting behavior. Its LLM installation supports Linux with vLLM, with a separate CPU-oriented option available. Apache-2.0 licensing makes the code free to use, while compute and external services remain the user’s responsibility. Arena is a separate hosted training and deployment platform; its accounts, credits and managed infrastructure are not included by installing the library.

What it can do

• Supervised fine-tuning and preference tuning for language models
• Multi-turn LLM agent reinforcement learning
• Single-agent, multi-agent, offline RL, and contextual bandit training
• Evolutionary hyperparameter optimization during training
• Local and distributed training, including multi-GPU options
• Train and evaluate agents with custom Gymnasium or PettingZoo environments

Works with

• PyTorch
• Hugging Face Transformers
• PEFT
• Datasets
• vLLM
• Gymnasium
• PettingZoo
• OpenEnv

Before you choose

• The vLLM-based LLM installation targets Linux; a separate CPU-oriented option omits vLLM.
• Provide suitable compute, data and dependencies for the chosen algorithm and workload.
• Arena is a separate hosted product with its own access and commercial terms.

Listed under

• Coding
• Agent training
• Developer tools
• Enterprise ML
• Fine-tuning
• Hyperparameter optimization
• Open-source models

Get started

1. Install `agilerl` for classic RL and the Arena CLI, or `agilerl[llm]` for LLM post-training.
2. Prepare a training manifest or configure a trainer with an environment, dataset, algorithm, and training settings.
3. Run training on suitable local or customer-controlled compute and review checkpoints and evaluation metrics.

The Python library is open source under Apache-2.0. Operating costs depend on compute, data and external services. Arena hosted accounts and credits are separate.

Source: github.com/AgileRL/AgileRL, checked 2 Nov 2026. Prices can change; check before you buy.

Sources for this overview

Product pages, documentation, and pricing checked 2 Nov 2026. Each source supports part of the overview; availability and prices can change.

• GitHub - AgileRL/AgileRL: The complete post-training stack for LLMs, plus RL for any problem. SFT, DPO and multi-turn agentic RL, with evolutionary hyperparameter optimization built in. · GitHub ↗github.com
• AgileRL | Specialized agents for your task ↗agilerl.com
• About AgileRL | A specialist agent for every enterprise task ↗agilerl.com
• AgileRL | Specialized agents for your task ↗agilerl.com
• Arena Documentation ↗docs.arena.agilerl.com

Company website

github.com: product, pricing and policy pages

YC directory mirror

Batch and company name

Social accounts

Accounts the company links, and its posts on X

Human review

Checked by a person

knowledge_base

manual_review

Last checked 2 Nov 2026. Every field keeps its source and date; unknown means we couldn’t confirm it.