Skip to content

Latest commit

 

History

5 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

Awesome Agent RL Environments Awesome

A curated list of training and evaluation environments for LLM / VLM agents, with a focus on the era of multi-turn reinforcement learning (post-DeepSeek-R1, post-GPT-5-thinking).

If you train, evaluate, or build LLM agents, you've probably noticed the explosion of "Gym for Agents"-style projects in 2024–2026 — SWE-Gym, GEM, RAGEN, VAGEN, AgentGym, verifiers, SkyRL, ART, prime-rl, WebRL, OSWorld, AndroidWorld, AppWorld, TheAgentCompany… Each of them ships its own task suite, its own RL trainer, its own evaluation harness. This list is here to keep track.

Scope. We only include projects that satisfy at least one of:

  1. Provide an interactive environment where an LLM/VLM agent issues actions, receives observations, and gets a reward / verifiable signal.
  2. Are explicitly designed for training agents via RL (PPO, GRPO, RLOO, DPO-on-trajectories, etc.) — not just evaluation.
  3. Are a gym-style framework that wraps multiple environments behind a unified API.

We intentionally exclude pure Atari / MuJoCo / classic RL benchmarks, and we exclude single-turn evaluation benchmarks (e.g., MMLU, GSM8K) — those are well-covered elsewhere.

Every entry below has a verified GitHub link. PRs that add a missing project, fix a broken link, or report a benchmark number are all appreciated. See CONTRIBUTING.md.


Contents


TL;DR — pick the right framework

Framework Best for RL algorithms Native envs Repo
GEM Math, code, QA, games — anything you want a gym-style API for PPO, GRPO, RLOO Math, Code, Reasoning, QA, Games axon-rl/gem
RAGEN Multi-turn reasoning agents in stochastic envs StarPO (custom) Sokoban, FrozenLake, WebShop, Lean, Sudoku, …(10 envs) RAGEN-AI/RAGEN
VAGEN Vision-Language-Model agents with world-model RL PPO + World-Modeling RL Visual multi-turn tasks RAGEN-AI/VAGEN
AgentGym / -RL Broadest task coverage (14 envs across web, tools, games, embodied) PPO, GRPO, RLOO, REINFORCE++ 14 envs, 89 tasks WooooDyy/AgentGym-RL
verifiers Trainer-agnostic env wrappers + reward verifiers bring your own (SkyRL, Tinker, …) env registry growing on Environments Hub PrimeIntellect-ai/verifiers
SkyRL Long-horizon, real-world agent training at scale full-stack (train + tx + agent) plugs into verifiers envs NovaSky-AI/SkyRL
prime-rl End-to-end SFT + RL + evals with verifiers integration PPO, GRPO SWE & agentic envs PrimeIntellect-ai/prime-rl
OpenPipe ART Single-GPU GRPO with ergonomic Python harness GRPO bring your own task OpenPipe/ART
verl-agent Extending veRL for agentic RL with GiGPO Group-in-Group PO WebShop, ALFWorld, AppWorld, … langfengQ/verl-agent
SWE-Gym Specifically training SWE agents on real GitHub issues any (PPO/GRPO/SFT) 2,438 Python tasks SWE-Gym/SWE-Gym

General RL-for-LLM Gyms (cross-task frameworks)

Frameworks that wrap many environments behind a unified API, so you can train once and roll out everywhere.

  • GEM: A Gym for Agentic LLMs (ICLR 2026) — General Experience Maker for LLMs. OpenAI-Gym-style API spanning math, code, general reasoning, QA, and games (Mastermind, Minesweeper, Hangman, …). 📄 Paper

  • RAGEN — Reasoning AGENt. Built around StarPO (State-Thinking-Actions-Reward Policy Optimization). 10 built-in envs: Sokoban, FrozenLake, WebShop, DeepCoder, SearchQA, Lean, Bandit, Countdown, MetaMathQA, Sudoku.

  • VAGEN — RAGEN's sibling for VLM agents. Proposes World-Modeling RL: explicitly supervises a VLM's world-model reasoning during multi-turn rollouts.

  • AgentGym (ACL 2025) — 14 environments across web navigation, text games, household tasks, digital games, embodied tasks, tool use, programming. Real-time concurrent agent interaction.

  • AgentGym-RL — Multi-turn RL training extension of AgentGym. Supports PPO, GRPO, RLOO, REINFORCE++. 📄 Paper

  • verifiers — Originally by @willccbb, now stewarded by Prime Intellect. Trainer-agnostic environment wrappers + reward verifier protocols. Supported as a backend by SkyRL and Tinker.

  • SkyRL (NovaSky-AI) — Full-stack modular RL library: skyrl-train (training framework), skyrl-tx (Tinker-API backend), skyrl-agent (long-horizon agent layer). 📄 SkyRL-Agent paper

  • prime-rl — Agentic RL training at scale. End-to-end SFT + RL + evals with native verifiers integration.

  • OpenPipe ART (Agent Reinforcement Trainer) — Ergonomic Python harness for GRPO on Qwen3, GPT-OSS, Llama, and more. Designed for maximum efficiency on a single GPU; builds on Unsloth.

  • verl-agent — Extends veRL for LLM/VLM agent RL. Official code for Group-in-Group Policy Optimization (GiGPO). Includes WebShop, ALFWorld, AppWorld, and other interactive envs.

  • AgentLab (ServiceNow) — Open-source framework for developing, testing, and benchmarking web agents on diverse tasks. Used together with BrowserGym.


Coding & Repository Environments

Environments where the agent edits real codebases, runs tests, and gets binary pass/fail rewards.

  • SWE-Gym (ICML 2025) — The first public training environment combining real-world GitHub issues with pre-installed dependencies and executable unit tests. 2,438 Python tasks. SWE-agents trained on SWE-Gym reach 32.0% on SWE-Bench Verified (open-weights SOTA at release). 📄 Paper

  • R2E-Gym (COLM 2025) — Procedural environment generation + hybrid verifiers for scaling open-weights SWE agents. Removes the bottleneck of hand-curated tasks. 📄 Paper

  • SWE-Bench — The original "can your agent fix a real GitHub issue?" benchmark; its Verified split (500 human-validated tasks) is the de-facto standard. 📄 Paper

  • Multi-SWE-Bench (NeurIPS 2025 D&B) — SWE-Bench across 7 languages (Java, TypeScript, JavaScript, Go, Rust, C, C++). 1,632 expert-annotated instances. 📄 Paper

  • SWE-Bench-Live (NeurIPS 2025 D&B, Microsoft) — Continuously updated SWE-Bench. 1,319 tasks from real GitHub issues since 2024 across 93 repos; refreshed monthly via the RepoLaunch pipeline to prevent contamination. 📄 Paper

  • SWE-rebench (Nebius) — Continuously updated decontaminated SWE benchmark; 21,000+ Python tasks suitable for RL of SWE agents. 📄 Paper

  • SWE-Bench Pro (Scale AI) — Enterprise-grade long-horizon tasks: 1,865 problems from 41 actively maintained repos. Top models score ~23% (vs. ~70% on Verified). 📄 Paper


Web / Browser Environments

Self-hosted realistic websites where agents must navigate, click, type, and complete multi-step tasks.

  • WebArena — The original realistic, reproducible web environment. Self-hosted instances of GitLab, Shopping, Reddit, Map, WikiCFM. 📄 Paper

  • VisualWebArena — Multimodal extension: 910 visual web tasks over Classifieds, Shopping, Reddit. 📄 Paper

  • BrowserGym (ServiceNow) — Unified Gym wrapper for WebArena, VisualWebArena, MiniWoB++, AssistantBench, and WorkArena. The de-facto standard for plugging your web agent into many envs at once.

  • WebArena-Verified (ServiceNow) — Verified, version-controlled WebArena tasks with deterministic evaluators. Reproducible benchmarking.

  • Mind2Web (NeurIPS 2023 Spotlight) — First generalist web-agent dataset/benchmark; later spawned Mind2Web 2 and Mind2Web Live for execution-based evaluation.

  • WebRL (THUDM) — Self-evolving online curriculum RL for web agents. Llama-3.1-70B trained with WebRL hits 49.1% on WebArena-Lite, beating GPT-4-Turbo (17.6%). 📄 Paper

  • AssistantBench — 214 realistic, time-consuming web tasks ("which gyms near me have classes on weekends before 7AM?"). State-of-the-art agents at release scored near zero. 📄 Paper

  • WorkArena (ServiceNow) — Enterprise SaaS workflows on the ServiceNow platform; WorkArena-L1 has 19,912 instances of 33 atomic tasks; WorkArena++ adds 682 compositional tasks.

  • ClawBench (arXiv 2026) — Containerized live-web environment with 153 write-heavy tasks, final-request interception, and trace-grounded binary evaluation for browser agents. 📄 Paper


Operating System / Desktop / Mobile Environments

The agent controls a real OS (Ubuntu, Windows, macOS, Android) and must operate apps the way humans do.

  • OSWorld (NeurIPS 2024) — First-of-its-kind scalable real-computer environment. 369 desktop tasks across Ubuntu/Windows/macOS spanning file I/O, multi-app workflows, OS settings. Humans 72.4% vs. best agent ~12.2% at release. 📄 Paper

  • AndroidWorld (Google Research, ICLR 2025) — Dynamic Android benchmark: 116 hand-crafted tasks across 20 apps, parameterized into millions of variations. Best agent at release ~30.6%. 📄 Paper

  • Windows Agent Arena (Microsoft) — 150+ Windows-specific tasks adapted from OSWorld. Real Windows OS with the full app ecosystem.

  • Agent-S (Simular AI) — Open agentic framework that "uses computers like a human." Strong reference for production-style computer-use stacks.


Tool-Use & Multi-Tool Environments

Agents call external APIs / tools and chain them together. Reward is usually a verifier checking the final state or output.

  • ToolBench (ICLR 2024 Spotlight, OpenBMB) — 16,000+ real-world APIs over 3,451 tools. Both training data and evaluation harness; the canonical tool-learning corpus.

  • AppWorld (ACL 2024 Best Resource, Stony Brook NLP) — Controllable simulated world of 9 apps (Amazon, Spotify, …) + 457 APIs + 100 simulated users. 750 day-to-day agent tasks. First-class MCP support.

  • τ-bench (Sierra) — Tool-Agent-User interaction in airline and retail customer-service domains. Measures consistency (success rate across 8 repeated runs), not just one-shot.

  • τ²-bench (Sierra) — Updated τ-bench with a banking domain, voice evaluation modality, and fixes to airline / retail tasks. Current recommended version.

  • ToolSandbox (Apple) — Stateful, conversational, interactive tool-use eval. Implicit state dependencies between tools + on-policy user simulator + dynamic milestone evaluation. 📄 Paper

  • MINT-Bench (ICLR 2024) — Multi-turn interaction with tools + natural-language feedback. 586 representative instances repurposed from 8 datasets across reasoning, code, decision-making. 📄 Paper

  • AgentBoard (NeurIPS 2024 Oral, HKUST-NLP) — Analytical evaluation board: 9 tasks, 1,013 environments. Fine-grained progress / sub-skill / trajectory inspection beyond final success rate.

  • TheAgentCompany (CMU) — Self-contained simulated software company (internal sites, data, employees). Best agent ~30% on consequential business tasks. 📄 Paper


Scientific & Research Environments

Long-horizon tasks where the agent has to reason over papers, run experiments, and produce science-grade output.

  • ScienceWorld (Allen AI) — Classic text-based science-experiment env: 10 interconnected locations, ~200 simulated lab objects. Still a tractable testbed for tool use + reasoning.

  • DiscoveryWorld (Allen AI) — Virtual environment for developing and evaluating automated scientific-discovery agents. Successor in spirit to ScienceWorld for the agent era. 📄 Paper

  • AI Scientist-v2 (Sakana AI) — Workshop-level automated scientific discovery via agentic tree search. Produced the first fully AI-generated paper to pass workshop peer review (ICLR 2025). 📄 Paper

  • CodeScientist (Allen AI) — Automated scientific discovery system for code-based experiments. End-to-end pipeline from hypothesis to executed experiment.

  • MLE-Bench (ICLR 2025, OpenAI) — 75 Kaggle-style ML engineering competitions with human Kaggle leaderboards as baselines. The standard ML-engineer agent benchmark. 📄 Paper

  • PaperBench (ICML 2025, OpenAI) — Replicate 20 ICML 2024 Spotlight/Oral papers from scratch, evaluated by 8,316 author-validated rubric sub-tasks. Best agent at release ~21%. 📄 Paper

  • ResearcherBench (GAIR-NLP) — 65 expert-curated frontier AI research questions across 35 subjects; dual rubric + factual evaluation for Deep Research systems. 📄 Paper

  • DRBench (ServiceNow) — Realistic enterprise deep-research benchmark.

  • LiveResearchBench (Salesforce AI Research) — Live benchmark for open-ended deep research "in the wild."

  • DeepResearch Bench — Comprehensive benchmark for deep-research agents.


Game & Simulation Environments

Constrained, verifier-friendly games used as RL playgrounds for LLMs. Several cross-task gyms above (notably GEM and RAGEN) also bundle game environments — see General RL-for-LLM Gyms.

  • TextWorld (Microsoft) — Sandbox text-game environment with controllable generation; still widely used to probe curriculum learning, generalization, and transfer in LLM agents. 📄 Paper

  • ALFWorld — Aligns text-based TextWorld policies with the visual ALFRED benchmark. Standard baseline for embodied / household agents. 📄 Paper

  • Crafter (Danijar Hafner) — Minecraft-inspired 2D open-world survival benchmark. Gym API; popular as a tractable open-ended task in LLM-agent papers.


Safety / Adversarial Environments

Environments specifically built to stress-test agent safety, prompt injection, and reward hacking.

  • AgentDojo (NeurIPS 2024, ETH Zürich SPyLab) — Dynamic environment to evaluate prompt-injection attacks and defenses for tool-using LLM agents. Used by US/UK AI Safety Institutes to stress-test Claude. 📄 Paper

  • InjecAgent (ACL 2024 Findings, UIUC) — Benchmark for indirect prompt injection via tool outputs. 1,054 test cases / 17 user tools / 62 attacker tools. 📄 Paper

  • ST-WebAgentBench — Safety + trustworthiness for web agents. 222 tasks paired with 646 policy instances across 6 ST dimensions (consent, boundary, hierarchy, robustness, error handling). 📄 Paper

  • ToolEmu (ICLR 2024 Spotlight) — LM-based emulation of tool execution with an automatic safety evaluator. 36 toolkits / 311 tools / 144 test cases. Originally framed as a risk-identification sandbox that lets you stress-test agent failures without real APIs.


Surveys & Related Reading


How to Contribute

We love PRs. To keep quality high, please:

  1. One project per PR. Smaller diffs review faster.
  2. Verify the GitHub link works before submitting — every entry in this list has a confirmed, live repo.
  3. Use this row format in the appropriate section:
    - **[Project Name](https://github.com/org/repo)** *(Venue Year, if any)* — One-sentence description that says what's distinctive. 📄 [Paper](https://arxiv.org/abs/XXXX)
    
  4. No closed, paywalled, or unavailable resources unless they are uniquely important.
  5. No generic LLM frameworks that aren't environments for agents.

See CONTRIBUTING.md for full guidelines. Open an issue with the label suggest-env if you're unsure whether something fits.


License

CC0 1.0 — Public domain. Use it any way you like.

Star History

If this list saves you time, please ⭐ the repo. It's the cheapest way to say thanks and helps others find it.

About

A curated list of training & evaluation environments for LLM/VLM agents (SWE-Gym, GEM, RAGEN, AgentGym, WebArena, OSWorld, ToolBench…). Updated weekly.

Topics

Resources

Contributing

Stars

7 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors