AI Agents
adk-python
Google's code-first toolkit for building real AI agents.
★ 21.7k starsApache-2.0
View guide →SWE-Gym: the first training environment for real-world software engineering agents and verifiers (ICML 2025).
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.
Decide your goal: training an agent, training a verifier for inference-time scaling, or reproducing the paper's results.
Clone the repository and review the scaffold-specific instructions for OpenHands or MoatlessTools.
Access the dataset from Hugging Face and pull the pre-built Docker image for the task instances you will train on.
Generate agent trajectories, fine-tune or rejection-sample your model, and evaluate on SWE-Bench splits using the included verifier approach.
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.
Documented generation of agent-environment interaction trajectories with OpenHands for fine-tuning models as SWE agents.
Verifiers trained on trajectories select the most promising solution among samples, improving results through best-of-n selection.
Reproduction instructions cover both the OpenHands and MoatlessTools agent scaffolds.
Docker images for each task instance are published on Docker Hub, removing environment setup friction.
The project reports open state-of-the-art results for its time on SWE-Bench Verified and Lite, with scaling analyses in the paper.
Start from the code and documentation for the training and reproduction workflows.
git clone https://github.com/SWE-Gym/SWE-Gym.git
cd SWE-GymDownload 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)Reproduce training or inference-time scaling with the documented scaffold instructions.
# see docs/OpenHands.md and docs/MoatlessTools.md in the repositorySWE-Gym was created by researchers from UC Berkeley, UIUC, CMU, and Apple, and is maintained under the SWE-Gym GitHub organisation.
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.
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.
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.
Clone the repo, download the dataset from Hugging Face, and follow docs/OpenHands.md or docs/MoatlessTools.md for the workflow you want.
Google's code-first toolkit for building real AI agents.
Benchmark computer-use agents on long, real-world desktop tasks.
Hugging Face's LeRobot: datasets, models and tools to make real-world robotics AI accessible
Team chat where AI agents are first-class teammates, not bots.
Loop Engineering — practical patterns, starters and a CLI for designing the loops that prompt and orchestrate AI coding agents, with a Loop Ready score
A code-first Java toolkit for building AI agents.
The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if SWE-Gym is not the right fit.
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
Free and open — grab the source, star the repo, and start creating.