Our take
Activeregular commits and releases- Activity
- 67
- Community
- 65
- Issues
- 60
- Pull requests
- 61
+54 stars in the last 30 days
- Merges more pull requests than 87% of the projects we track
Uni-Agent is an actively maintained, Apache-licensed framework for teams that need a shared path from agent execution to scalable evaluation and RL training, but it is still a young project.
Good fit if
- You are building long-horizon agents and want to reuse existing harnesses or compose agents from explicit task and tool abstractions.
- Your workflow needs concurrent inference, trace collection, and reinforcement learning with verl integration.
- You can work in Python and provide the sandbox and compute infrastructure your workloads require.
Look elsewhere if
- You need a turnkey agent product rather than a framework for building and training agents.
- Your project does not use Python or cannot accommodate the supported Python version range.
- You need a mature, long-established project with a longer operating history.
All health signals
| Last commit | 2026-10-25 (today) |
|---|---|
| Commits, last 90 days | 100+ |
| Releases, last 12 months | 2 (latest v0.1.0, 2026-09-23) |
| Contributors | 33 (top contributor: 45% of commits) |
| Issues closed, last 90 days | 4+ (typically closed in 2 days) |
| Pull requests merged, last 90 days | 81 (typically merged in 1 day) |
| Project age | 9 months |
Checked on 2026-10-26 with the GitHub API.
Project history
Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.
Stars
725Forks
156Change since Nov 14, 2026: +1 stars·0 forks
Overview
Uni-Agent is a Python framework for training long-horizon agents, integrating with verl for reinforcement learning. It addresses the challenge of running stateful agents across varied tasks and collecting their interactions as data that can be used for evaluation, supervised fine-tuning, and RL.
Teams can connect existing harnesses through its Gateway or build agents using reusable agent, tool, task, and sandbox abstractions. Its runtime supports parallel sessions and asynchronous training, with recipes and benchmark results available in the repository and documentation.
Key Features
- Connect existing harnesses, including Claude Code and Mini-SWE-Agent, through OpenAI- or Anthropic-compatible endpoints.
- Compose white-box agents from reusable
Agent,Tool,Task, andSandboxabstractions. - Run concurrent, stateful sessions with distributed workers, pooled Gateway sessions, and asynchronous scheduling.
- Associate trajectories, logs, and rewards with their originating sessions for evaluation and training.
- Generate training-ready interaction data for supervised fine-tuning and reinforcement learning.
- Use the same interaction stack for inference and RL training, with recipes for GRPO/GSPO-style objectives and partial rollouts.
- Run Harbor tasks through the inference pipeline as an additional task format.
Use Cases
- Agent researchers can test whether reinforcement learning improves a coding agent on software-engineering benchmarks.
- Teams with an existing agent harness can route its model requests through the Gateway to collect trajectories without replacing the harness.
- Training engineers can run asynchronous agent rollouts and use the resulting traces and rewards in RL workflows.
- Evaluation teams can run parallel sessions across tasks and compare agent implementations using the shared inference pipeline.
What you need
Detected in the repository
- Python >=3.10,<3.13 (from pyproject.toml)
- A test suite and automated checks on GitHub Actions
License in plain words
Apache-2.0permissive
- Commercial use: yes
- Modify and redistribute: yes
- You must keep: the license, the NOTICE file and a note of your changes
- Share your changes: no
- Includes an explicit patent grant from the contributors:
A summary, not legal advice: the LICENSE file is what applies.
Getting Started
Install Uni-Agent by following the installation guide. Then use the quickstart guides to launch a sandbox, run inference, or train an agent. See the README for examples and reproducible recipes.
Alternatives
- envharness: envharness composes layers around frozen interactive environments to target agent weaknesses, while uni-agent connects existing agent harnesses to RL training with verl.
| Project | Language | License | Stars | Status |
|---|---|---|---|---|
| uni-agent | Python | Apache-2.0 | 693 | Active |
| envharness | Python | Apache-2.0 | 679 | New |
Considerations
Uni-Agent is young, although its recent commit and release activity is regular. It has contributions from a broad group, rather than depending mostly on one contributor; recent pull requests are merged quickly, while issue closure activity is more limited. These signals are useful context, but do not replace evaluating the framework against your own training workload.
The framework's large-scale workflows rely on sandbox backends and distributed execution infrastructure. The README acknowledges remote sandbox services such as veFaaS and Modal, and one training recipe uses H100 GPUs; these are examples of deployment needs, not stated universal requirements. The project requires Python 3.10 through 3.12. Its Apache-2.0 license permits broad use, subject to the license's terms.
Found this useful?
Share it with someone who would like uni-agent.