Uni-Agent: A Scalable Framework for Training Long-Horizon AI Agents

Use at your own risk

OSRepos shares public repositories for knowledge and discovery only. Any installation, execution, configuration, or use of code from these repositories is the user's own responsibility. Always review the repository, source code, dependencies, licenses, and security implications before running or installing anything. OSRepos is not responsible for issues, damages, or losses resulting from third-party repositories.

Introduction

Uni-Agent is a powerful Python framework designed for training long-horizon agents at scale. It allows you to bring any existing agent harness into reinforcement learning, unify diverse agent tasks through one extensible interface, and run agents concurrently at scale. The framework collects traceable trajectories as training-ready data for Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL).

Why Use It & Key Benefits

Uni-Agent offers several compelling features for developing and training advanced AI agents:

Plug in any agent harness: Connect harnesses like Claude Code and Mini-SWE-Agent, or any harness compatible with OpenAI or Anthropic model endpoints, directly to the Uni-Agent Gateway.
Decouple agents, tasks, and infrastructure: Build white-box agents from reusable Agent, Tool, Task, and Sandbox abstractions. Customize agent logic, tools, task environments, sandbox backends, and rewards independently while reusing the same evaluation and training runtime.
Run thousands of sessions concurrently: Execute over 1,000 long-horizon, stateful sessions with distributed workers, pooled Gateway sessions, isolated sandboxes, and asynchronous scheduling. Every trajectory, log, and reward remains associated with the correct session for reliable evaluation, RL training, and data synthesis.
Reproducible training, verifiable results: The project publishes runnable recipes with complete configurations, benchmark settings, result tables, and learning curves. These recipes provide tested starting points and make reported improvements easier to reproduce and verify.

The framework demonstrates strong performance in both parallel inference and verification, as well as agent reinforcement learning, with impressive scores on benchmarks like SWE-Bench and Terminal-Bench.