Introducing Uni-Agent: veRL’s Open‑Source Unified Framework for General‑Purpose Agent Training

Background

Open‑source agent projects such as OpenClaw have revealed a demand for infrastructure that can handle complex, general‑purpose scenarios, scale to large workloads, and integrate naturally with training pipelines. Existing frameworks perform well on benchmarks and single‑task demos but lack system‑level robustness for long‑running, real‑world applications.

Uni‑Agent Overview

Uni‑Agent (GitHub repository: https://github.com/verl-project/uni-agent) is a unified framework that spans the Build, Run, and Train stages of agent development. It aims to simplify agent construction, provide stable large‑scale execution, and enable continuous reinforcement‑learning evolution.

Design Principles and Core Modules

model : inference and decision‑making; supports external API services and self‑hosted back‑ends such as vLLM and SGLang.

tool : perception and action; serves as a plug‑in point for task‑specific functionality.

env : execution environment and state storage.

The modularity allows developers to add new capabilities with minimal code changes. For example, the arXiv search and recommendation tutorial adds only a new tool component (tutorial URL: https://uni-agent.readthedocs.io/en/latest/start/search_agent.html).

Run Layer – Scalable Execution

Uni‑Agent integrates Volcano Engine veFaaS Sandbox to provide:

Security : MicroVM isolation for each task, protecting against unknown code execution.

Performance : Image pre‑warming, resource pooling, and optimized scheduling that remain stable under ten‑thousand‑level concurrent tasks.

Scenario Adaptation : Support for Code, Browser, and Computer environments and custom images, enabling integration with real toolchains.

A lightweight live dashboard offers real‑time monitoring of task status, logs, and overall progress.

Train Layer – Continuous Evolution

Uni‑Agent couples with the verl reinforcement‑learning engine (GitHub: https://github.com/verl-project/verl) to support state‑of‑the‑art training techniques. Experiments on a large‑scale coding‑agent task used the open‑source R2E‑Gym dataset (~4,500 samples) to train a Qwen3‑Coder‑30B model. Results showed steady reward growth, overall validation improvement, and emergent capabilities, demonstrating that the Uni‑Agent training pipeline can continuously boost model performance in realistic tasks.

Analysis of rollout data revealed a pronounced long‑tail effect: wide variance in rollout length and execution time across samples. To address this, Uni‑Agent implements fully async and partial rollout strategies. Benchmarks indicate multiple‑fold efficiency gains over synchronous training while maintaining stable outcomes.

Long‑Term Vision

The goal is to move beyond agents that merely chat or call tools toward systems that can perceive, act, explore, and evolve within complex environments. Uni‑Agent is positioned as a foundational platform that provides unified, extensible, and scalable architecture to support such capabilities.