菜单
Build scalable LLM agent frameworks with automated generation of tools, prompts, and configurations, plus continuous optimization through experience accumulation and reinforcement learning—achieving 71.47% on WebWalkerQA and 72.8% on GAIA.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill youtu-agent-framework命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | youtu-agent-framework |
| title | Youtu-Agent: Scaling Agent Productivity with Automated Generation and Hybrid Policy Optimization |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2512.24615 |
| keywords | ["Agent Framework","Automated Agent Generation","LLM Agents","Policy Optimization","Agent RL"] |
| description | Build scalable LLM agent frameworks with automated generation of tools, prompts, and configurations, plus continuous optimization through experience accumulation and reinforcement learning—achieving 71.47% on WebWalkerQA and 72.8% on GAIA. |
Youtu-Agent is a modular framework for constructing and continuously improving LLM-based agents. It addresses two critical bottlenecks: the high manual effort required for agent configuration (tool integration, prompt engineering) and the difficulty in adapting deployed agents without expensive fine-tuning.
Key Innovation: Automated generation of complete agent configurations from high-level task descriptions, combined with two optimization mechanisms: Agent Practice (low-cost in-context learning) and Agent RL (end-to-end reinforcement learning at scale).
The framework organizes agent execution into three hierarchical layers:
Environment Layer Provides foundational execution context and primitives. Typical backends include browser automation (Playwright), OS shells, and sandboxed code execution (E2B).
Tools Layer Encapsulates atomic and composite operations: (1) environment-related tools wrapping low-level APIs, (2) environment-independent utilities, (3) MCP tools for external services.
Agent Layer Houses the LLM-driven planner with a perceive-reason-act loop and Context Manager for managing long-horizon interactions.
A deterministic four-stage pipeline for routine tasks:
This mode achieves 100% configuration validity on the 80-task AgentGen benchmark.
For complex or ambiguous requirements, deploy an Architect Agent with tools: search_tool, create_tool, ask_user, and create_agent_config. The meta-agent dynamically plans generation through multi-turn clarification and tool synthesis.
Example: Given "Summarize today's trending papers on multi-agent systems and download PDFs," the meta-agent:
search_tool to find existing arxiv toolkitcreate_tool to synthesize fetch_daily_papers toolcreate_agent_config to assemble final configurationTool synthesis implementation:
def fetch_daily_papers(date: str) -> str:
"""Crawl daily papers from aggregation site.
Args:
date (str): date in format YYYY-MM-DD
"""
papers = list_daily_papers(date=date)
return "\n".join([f"{asdict(paper)}" for paper in papers])
Meta-Agent mode achieves 98.75% configuration validity and 68.75% end-to-end task completion on AgentGen-80.
The Agent Practice module enables low-cost agent improvement without parameter updates through Training-free Group Relative Policy Optimization (Training-free GRPO).
Mechanism:
Results:
Training dynamics show steady performance improvement with decreasing tool usage, indicating more efficient problem-solving strategies.
For applications requiring significant lasting improvement, Agent RL provides end-to-end reinforcement learning at production scale.
Scalability Solutions:
Stability Solutions:
Results on Qwen2.5-7B:
YAML-based configuration system enables both manual composition and automated generation:
agent:
name: research_agent
instructions: "You are a helpful research assistant..."
env:
name: e2b
config: {}
context_manager:
name: base
config: {}
toolkits:
search:
activated_tools: ["search", "web_qa"]
python_executor:
activated_tools: ["execute_python_code"]
The standardized format serves as the target schema for automated generation and facilitates sharing of agent variants across teams.
WebWalkerQA (680 questions):
GAIA Text-only Subset (466 questions):
Use Youtu-Agent when:
When NOT to use:
Full implementation available at: https://github.com/TencentCloudADP/youtu-agent
基于 SOC 职业分类
