AI Agents
adk-python
Google's code-first toolkit for building real AI agents.
★ 21.7k starsApache-2.0
View guide →// agents · SWE-Gym/SWE-Gym
SWE-Gym
SWE-Gym is an open-source ai agents project by SWE-Gym (SWE-Gym/SWE-Gym on GitHub). It has 748 GitHub stars and 46 forks, is released under the Apache-2.0 license, and was last updated on GitHub on 29 July 2025.
SWE-Gym: the first training environment for real-world software engineering agents and verifiers (ICML 2025).
★ 748 stars⑂ 46 forksApache-2.0updated 2025-07
▦ visual — this project's README ships no embeddable screenshot,
so here's a placeholder. The guide below is complete regardless.
so here's a placeholder. The guide below is complete regardless.
// project facts
Project facts
- Repository
- SWE-Gym/SWE-Gym
- GitHub stars
- 748
- Forks
- 46
- License
- Apache-2.0
- Last updated on GitHub
- 29 July 2025
- Category
- AI Agents
- Listing
- Full install guide
- Official site
- Official site
// how it works
How it works
- Clarify your goal
Decide your goal: training an agent, training a verifier for inference-time scaling, or reproducing the paper's results.
- Get the project
Clone the repository and review the scaffold-specific instructions for OpenHands or MoatlessTools.
- Install and configure
Access the dataset from Hugging Face and pull the pre-built Docker image for the task instances you will train on.
- Run your first task
Generate agent trajectories, fine-tune or rejection-sample your model, and evaluate on SWE-Bench splits using the included verifier approach.
// features
Why people love it
Real tasks with verification
Thousands of real-world tasks drawn from 11 Python repositories, each with repository context, an executable environment, and test-based verification, plus a smaller Lite split.
Agent training trajectories
Documented generation of agent-environment interaction trajectories with OpenHands for fine-tuning models as SWE agents.
Verifier training for inference-time scaling
Verifiers trained on trajectories select the most promising solution among samples, improving results through best-of-n selection.
Two scaffold paths
Reproduction instructions cover both the OpenHands and MoatlessTools agent scaffolds.
Pre-built instance images
Docker images for each task instance are published on Docker Hub, removing environment setup friction.
Published open baselines
The project reports open state-of-the-art results for its time on SWE-Bench Verified and Lite, with scaling analyses in the paper.
Advertisement
// run it
Run it yourself
- Clone the repository
Start from the code and documentation for the training and reproduction workflows.
git clone https://github.com/SWE-Gym/SWE-Gym.git cd SWE-Gym - Get the dataset and environment images
Download the dataset and models from the project's Hugging Face page and pull instance images from Docker Hub. # see README reproducing results section
# dataset and models: https://huggingface.co/SWE-Gym (see README) - Follow a scaffold guide
Reproduce training or inference-time scaling with the documented scaffold instructions.
# see docs/OpenHands.md and docs/MoatlessTools.md in the repository
// use cases
What you can build
- Training an open model to resolve real GitHub issues more reliably
- Learning a verifier that picks the best of several candidate patches
- Studying how agent performance scales with training and inference compute
- Bootstrapping self-improvement loops where a model trains on its own verified trajectories
Advertisement
// faq
Questions, answered
Who maintains it?
SWE-Gym was created by researchers from UC Berkeley, UIUC, CMU, and Apple, and is maintained under the SWE-Gym GitHub organisation.
Is it free to use?
Yes, the project is Apache-2.0 licensed and its dataset and models are public on Hugging Face; training itself requires substantial compute that you provide.
Does it need a paid API key?
No API key is required for the environment itself, though the documented trajectory generation used hosted models such as GPT-4o and Claude, which need those providers' access if you follow the same recipe.
How does it compare to a well-known alternative?
SWE-Bench is an evaluation benchmark; SWE-Gym is a training environment with executable tasks and verification designed to improve performance on benchmarks like SWE-Bench.
How do I get started?
Clone the repo, download the dataset from Hugging Face, and follow docs/OpenHands.md or docs/MoatlessTools.md for the workflow you want.
// keep exploring
Similar tools
AI Agents
OSWorld-V2
Benchmark computer-use agents on long, real-world desktop tasks.
★ 353 starsApache-2.0
View guide →AI Agents
lerobot
Hugging Face's LeRobot: datasets, models and tools to make real-world robotics AI accessible
★ 28.0k starsApache-2.0
View guide →AI Agents
cumora
Team chat where AI agents are first-class teammates, not bots.
★ 3.9k starsMIT
View guide →AI Agents
Loop Engineering
Loop Engineering — practical patterns, starters and a CLI for designing the loops that prompt and orchestrate AI coding agents, with a Loop Ready score
★ 11.6k starsMIT
View guide →AI Agents
tau2-bench
A benchmark for tool-agent-user interaction in real-world domains.
★ 2.2k starsMIT
View guide →// alternatives
Alternatives in AI Agents
The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if SWE-Gym is not the right fit.
- openclaw★ 391.6k
A personal AI that really does things — on any OS, with your data staying yours.
- superpowers★ 294.9k
An agentic skills framework & software development methodology that works.
- ECC★ 275.2k
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
- hermes-agent★ 252.0k
The agent that grows with you — a Python framework for adaptive AI agents.
- deepseek-harness★ 245.5k
DeepSeek's plugin-based agent harness where everything is a plugin.
// more in agents
More in AI Agents
Nearby AI Agents projects in the catalogue, closest in stars to SWE-Gym — keep exploring the shelf either side of it.
- modelscan★ 781
Scan AI models for serialization attacks before you load them.
- agentsight★ 728
See what your AI agents do — at the system level.
- bolna★ 786
Open-source conversational voice AI agents.
- awesome-ai-tools★ 592
A very large curated list of AI tools, updated for 2026.
- mcp★ 789
An MCP server that puts Hyperbrowser's browsers behind the protocol.
- fastagency★ 547
Take multi-agent AutoGen workflows to production, fast.
Get SWE-Gym
Free and open — grab the source, star the repo, and start creating.