uni-agent: Train Long-Horizon Agents at Scale

Overview

Uni-Agent provides a common framework for connecting agent implementations to tasks and sandbox environments, then using those interactions for evaluation and training. It addresses the challenge of turning stateful, multi-step agent runs into traceable trajectories that can support supervised fine-tuning and reinforcement learning.

The project is aimed at researchers and developers building or evaluating agents over extended tasks. It can accommodate existing compatible harnesses or custom agents, while reusing the same interaction stack across inference and training. Distributed execution and sandbox backends support workloads that need concurrent sessions.

Key Features

• Connect existing agent harnesses through the Uni-Agent Gateway using OpenAI- or Anthropic-compatible endpoints.
• Define reusable agents, tools, tasks, and sandbox backends through separate abstractions.
• Run concurrent, stateful sessions with distributed workers and asynchronous scheduling.
• Associate trajectories, logs, and rewards with their sessions for evaluation and training data.
• Use a shared interaction stack for inference, verification, and agent reinforcement learning.
• Run Harbor tasks through the inference pipeline as an additional task format.
• Follow published training recipes with configurations, benchmark settings, and reported results.

Use Cases

• Agent researchers can generate and inspect long-horizon task trajectories for evaluation or training.
• Teams with an existing compatible agent harness can connect it to a shared runtime rather than replacing its interaction layer.
• Developers experimenting with agent reinforcement learning can use the provided recipes and interaction stack to run training workflows.
• Benchmark maintainers can run agent implementations across tasks and sandbox environments, including Harbor-formatted tasks.

Getting Started

Start with the installation guide, then follow the documentation for sandbox setup, inference, and RL training. See the repository README for project examples and links to reproducible recipes.

Alternatives

• envharness: envharness composes layers around frozen benchmark environments, while uni-agent provides a broader stack for running and training long-horizon agents.

Considerations

• This is a framework for agent infrastructure and experimentation, not a ready-made agent or turnkey benchmark service.
• Running long-horizon sessions at scale involves distributed workers and isolated sandbox environments. The README describes local and cloud sandbox options, but does not provide a single universal deployment command.
• The documented RL path relies on verl features, including colocate_async; check the installation guide for compatibility and setup requirements.
• The repository reports benchmark and training results, but results depend on the listed agent, model, task, and settings. Consult the linked documentation before comparing them with other evaluations.