DOCSEARCHbyAlgolia
Submit Docs

Youtu-Agent

repository·main·Indexed 26 days ago

https://github.com/tencentcloudadp/youtu-agent

A flexible framework for building, running, and evaluating autonomous agents using open-source models. It features automated agent and tool generation, continuous experience learning via training-free GRPO, and scalable reinforcement learning. The framework supports various use cases including data analysis, literature review, RAG, and PPT generation, and provides a DBTracingProcessor system for analyzing tool calls and agent trajectories.

Tokens
26.2K
Snippets
78
Records
159
Agent score
89%

What's inside youtu-agent

  1. Overview of Youtu-Agent Architecture

    main

    Youtu-Agent (abbreviated as utu in code) is a modular framework for building, running, and evaluating autonomous agents. The architecture follows a clear separation of concerns:

    • AgentConfig: Defines the configuration for an Agent.
    • Agent: The core logic that operates within an Environment and utilizes Toolkits to perform actions.
    • Environment: Provides the state and context (e.g., ShellLocalEnv for filesystem access or BrowserEnv for web interaction).
    • Toolkits: Collections of tools (e.g., web search, file manipulation, code execution) that grant agents capabilities.
    • Evaluation Framework: A benchmarking system used to evaluate agent performance through data management, processing, and automated execution/judging.

  2. Overview of Youtu-Agent Toolkits

    main

    Toolkits are collections of related tools that an agent uses to perform actions. They extend an agent's capabilities. There are three types of toolkits:

    • builtin: Default toolkits provided by the framework.
    • mcp: Toolkits accessible via the Model Context Protocol.
    • customized: User-created toolkits.

    All builtin toolkits inherit from the AsyncBaseToolkit abstract base class.

  3. Overview of Youtu-Agent

    main

    Youtu-Agent is a flexible, high-performance framework designed for building, executing, and evaluating autonomous agents. It is optimized for open-source models (such as DeepSeek-V3) and supports advanced capabilities like data analysis, file processing, and deep learning.

    Key capabilities include:

    • Automated Agent Generation: Supports Workflow mode for standard tasks and Meta-Agent mode for complex requirements, with high success rates in synthesizing tools, prompts, and configurations.
    • Agent Practice: A module for continuous, low-cost evolution using Training-Free GRPO, allowing agents to accumulate experience via in-context optimization without parameter updates.
    • Agent RL: A scalable pipeline for end-to-end reinforcement learning that supports distributed training up to 128 GPUs.
    • Extensibility: Built upon openai-agents, it supports various model APIs (from DeepSeek to gpt-oss), tool integrations, and framework implementations.

  4. Overview of Agent Practice with Training-Free GRPO

    main
    The utu/practice module implements Agent Practice using Training-Free Group Relative Policy Optimization (GRPO). This approach enhances agent performance without updating LLM parameters by leveraging group relative semantic advantages and iteratively distilling high-quality experiential knowledge. It is a cost-effective solution for improving agent capabilities through experience.

  5. Explore Youtu-Agent Features and Use Cases

    main

    Youtu-Agent provides out-of-the-box support for several practical agentic tasks:

    • Data Analysis
    • Literature Review
    • Personal File Organization
    • Retrieval-Augmented Generation (RAG)
    • PPT Generation

    Additionally, the framework supports Agent Skills, which allow you to extend agents with modular, domain-specific knowledge and workflows.

  6. Core Concepts of Youtu-Agent

    main

    Understanding the fundamental building blocks of the framework:

    • Agent (智能体): An LLM configured with specific prompts, tools, and an environment.
    • Toolkit (工具包): A collection of encapsulated tools available for the Agent to use.
    • Environment (环境): The world in which the Agent operates (e.g., a browser or a shell).
    • ContextManager (上下文管理器): A configurable module used to manage the Agent's context window.
    • Benchmark (基准): An encapsulated workflow for a specific dataset, including preprocessing, execution, and judgment logic.

  7. Review ACL 2025 Outstanding Papers Research Themes

    main

    The ACL 2025 Outstanding Papers can be categorized into six primary research directions:

    • A. Diagnostics of language-model behavior: Focuses on theoretical lenses (e.g., Zipf-law reformulation, Capability Salience Vectors) to expose LLM strengths and weaknesses.
    • B. Dataset creation & resource expansion: Focuses on large, multilingual, or domain-specific corpora (e.g., synthetic instruction corpora, hallucination benchmarks, African/Indian/Middle-Eastern language resources).
    • C. Efficient model architectures & scaling: Focuses on reducing computational cost (e.g., byte-level patches, token-recycling, embedding standardization).
    • D. Evaluation & benchmarking advances: Focuses on new metrics and efficient benchmark pruning (e.g., MiniLongBench).
    • E. Safety, ethics, and societal impact: Focuses on plagiarism detection, hallucination taxonomies, and jailbreak safety gaps.
    • F. Multilingual & typologically diverse NLP: Focuses on methods for low-resource or typologically distinct languages (e.g., MoI-MoE architecture).

  8. Understand Agent Environments (Env)

    main

    An Environment (Env) represents the world in which the agent operates. It provides the agent with its current state and a set of tools for interaction. The framework uses a factory function, get_env, to instantiate the environment based on the agent's configuration file.

    All environments implement the following core interface:

    • get_state() -> str: Returns a string describing the current state (injected into the agent's prompt).
    • get_tools() -> list[Tool]: Returns a list of Tool objects available to the agent.
    • build() / cleanup(): Manages the lifecycle (starting/stopping services).

  9. Key features and architecture of Youtu-Agent

    main

    Youtu-Agent is designed for simplicity, modularity, and low cost. Key technical features include:

    • Built on openai-agents: Inherits streaming, tracing, and agent loop capabilities from the openai-agents SDK, ensuring compatibility with responses and chat.completions APIs (e.g., for gpt-oss models).
    • Fully Asynchronous: Designed for high performance and efficient execution, particularly beneficial for evaluation tasks.
    • YAML-based Configuration: Uses structured YAML files for manageable agent settings.
    • Tracing and Analysis: Supports OTEL and includes a DBTracingProcessor system (upcoming) for detailed analysis of tool calls and agent traces.

  10. Review Youtu-Agent core features

    main

    Youtu-Agent provides several key capabilities for agent development:

    • Built on openai-agents: Inherits streaming, tracing, and agent-loop capabilities, compatible with responses and chat.completions APIs.
    • Fully asynchronous: High-performance execution suitable for benchmarks.
    • Tracing & analysis: Includes a DBTracingProcessor system for in-depth analysis of tool calls and agent trajectories.
    • YAML-based configuration: Structured and manageable agent configurations.
    • Automation: Supports automatic agent generation and tool generation/optimization.

  11. How Agent Skills Work

    main

    When an agent with enabled_skills is initialized:

    1. Skill Deployment: Skill folders are copied from .agent/skills/ to the workspace .agent/skills/ directory.
    2. Prompt Injection: A system prompt is added to the agent's context listing all available skills and their descriptions.
    3. On-demand Reading: The agent invokes skills by using the openskills CLI tool to read the skill content via the command: openskills read <skill-name>.

  12. Understand Youtu-Agent core concepts

    main

    To build with Youtu-Agent, familiarize yourself with these fundamental building blocks:

    • Agent: An LLM configured with specific prompts, tools, and an environment.
    • Toolkit: An encapsulated set of tools that an agent can use.
    • Environment: The world in which the agent operates (e.g., a browser, a shell).
    • ContextManager: A configurable module for managing the agent's context window.
    • Benchmark: An encapsulated workflow for a specific dataset, including preprocessing, rollout, and judging logic.