Tencent open-sources a new intelligent agent framework: no training required, no recharge needed, achieving SOTA Agent with open-source models

Intelligent agents are becoming the key carriers for the implementation of large models. Whether for research exploration or practical applications, an efficient, user-friendly, and reproducible intelligent agent framework is the infrastructure that drives the development of the intelligent agent ecosystem.

However, researchers and developers often encounter some challenges: for example, high entry barriers, complex dependency environments, difficulties in reproducing experiments, and the need to train models or recharge expensive overseas closed-source model APIs to achieve reliable performance.

To address this, Tencent Youtu Lab has officially open-sourced Youtu-agent — a simple yet powerful intelligent agent framework that balances minimalist design with high performance, capable of serving both research baseline needs and supporting practical application construction. Most importantly, it achieves excellent performance without the need to train models or rely on closed-source APIs.

Youtu-agent has demonstrated leading performance on multiple challenging benchmarks:

• WebWalkerQA: Achieved 71.47% accuracy based on DeepSeek-V3.1, setting a new open-source SOTA;
• GAIA (text subset): Achieved 72.8% Pass@1 based on DeepSeek-V3 without needing to recharge closed-source models like Claude/GPT, validating its strong research and application potential.

This means that Youtu-agent, under fully open-source and reproducible conditions, is already approaching or even surpassing some intelligent agent frameworks that rely on paid tools.

Ready to Use

Overall, Youtu-agent has the following core highlights:

1. Open-source Friendly & Cost-sensitive

Completely based on the open-source ecosystem, it does not rely on closed models and is adaptable to accessible, low-cost deployment environments.

1. Flexible Architecture

Built on openai-agents, it is compatible with various model APIs and toolsets such as DeepSeek and gpt-oss.

1. Automatic Intelligent Agent Generation

Based on YAML configuration and a "meta-agent" dialogue mechanism, users only need to describe their needs to generate and run intelligent agent configurations with one click.

1. Concise and Efficient

Modular + asynchronous design supports streaming, tracing, and agent-loop, enabling efficient debugging and expansion.

More importantly, Youtu-agent is not a "paper framework" in the lab but a ready-to-use intelligent agent tool for real-world scenarios.

The research team has outlined four typical cases covering file management, data analysis, academic research, and broad reviews, showcasing the framework's flexibility and practicality Case 1: Local File Management

In the scenario where teaching assistants grade assignments, users only need to place the files submitted by students into a designated folder. Once the Youtu-agent is activated, it scans all files in sequence, first determining whether the format is PDF.

If it is a PDF, it will automatically rename it to the standard format of "Student ID - Name" according to the rules; if it is not a PDF, it will archive it into a separate folder. The entire process requires no human intervention, seamlessly transitioning from file recognition to naming and then to classification and archiving.

Case 2: Data Analysis

When faced with the cat_breeds_clean.csv file in the Kaggle dataset, users only need to provide the path. The Youtu-agent reads the CSV content, and after data cleaning and statistical analysis, it automatically generates structured conclusions and trends, transforming them into an intuitive HTML report.

In this process, the raw data is gradually transformed into high-value information: from reading tables, to statistical extraction, and finally to result visualization, forming a complete pipeline.

Case 3: Paper Analysis

In scientific research, researchers only need to input a PDF paper. The Youtu-agent first parses the document content and extracts the core parts; then it calls a search tool to externally retrieve related research work; finally, it integrates the paper interpretation with external results to generate a Markdown report.

The entire process acts like a "paper assistant": first understanding the paper, then searching for comparisons, and finally writing a clear and organized research note.

Case 4: Wide Research

When users input a broad topic, the Youtu-agent enters a composite mode. It first calls a search tool to collect a large amount of dispersed information; then it uses document processing tools for organization; finally, the agent analyzes and summarizes, generating a structured Markdown overview.

This process is equivalent to automating "research": from the proposal of needs, to the collection of massive data, and finally to the formation of reports, the entire chain is completed collaboratively by multiple sub-agents.

Design Principles: DITA

The researchers of the Youtu-agent proposed the DITA Principles, summarizing four key dimensions of agent design:

• Demand: Clearly define the task objectives, derived from the requirements in system prompts or user input intentions.
• I/O: Define the input forms of the task (such as CSV, PDF, str) and the output forms (such as HTML reports, Markdown).
• Tools: Select or generate appropriate tools based on the task, such as search tools and document processing tools.
• Agent Pattern: Determine the interaction mode of the agent, such as single agent, plan-and-execute, compound

Automated Agent Generation

In addition to the DITA framework, Youtu-agent also emphasizes a core highlight: Automated Agent Generation.

In traditional methods, users often need to manually write prompts, configure tools and parameters, which not only presents a high usage barrier for beginners but also requires experienced users to invest a significant amount of time. Youtu-agent's solution is to standardize and automate this process.

Specifically, it adopts a unified YAML configuration format to describe the agent's input and output, tool calls, and interaction paradigms in a structured manner; at the same time, it introduces a "meta-agent" that allows users to simply state their task requirements, and the system will clarify intentions through interaction with the user and automatically generate a complete configuration file. The generated configuration can be directly loaded and run, enabling one-click testing and usage.

In terms of user experience, users only need to run python scripts/gen_simple_agent.py first, and the meta-agent will generate the corresponding configuration based on the requirements; then, by running python scripts/cli_chat.py --stream --config generated/xxx, they can immediately start and test the generated agent. This mechanism significantly reduces the difficulty of customizing agents, allowing beginners to quickly get started while also helping advanced users save tedious configuration and debugging time.

Quick Start

For researchers/model trainers, Youtu-agent can serve as a stronger open-source baseline than ReAct, with one-click evaluation scripts to assist in experiments and ablation studies.

For application developers, Youtu-agent can act as a proven scaffold, allowing for lightweight entry to build real intelligent applications.

For AI enthusiasts, Youtu-agent offers rich examples and debuggability, making development and exploration intuitive and efficient.

So how do you use this agent framework? It's simple; Youtu-agent has open-sourced the complete code and examples on GitHub, and you can get started in just three steps:

Step 1: Get the Code

Enter the following in the command line:

Step 2: Run the simplest agent

Youtu-agent comes with built-in configuration templates, and you can define your agent's behavior by simply writing a YAML file. The project provides a basic agent template with a search tool located at configs/agents/default.yaml. You can run the following code for a one-click experience:

Step 3: Explore more examples

You can also try running more practical cases from the official examples. Some examples require the agent to have internet search capabilities, so you need to configure the tool API in the .env file in advance:

For example, if you want the agent to automatically search for information online and generate an SVG introduction image around the theme "DeepSeek V3.1 new features," you can directly run the following code:

If you want to preview the agent's operation visually using the web UI, you can download the packaged frontend UI files from the releases of Youtu-agent and install them locally:

Run the web version of the SVG image generation command:

For example, for the question "What are the highlights of deepseek-v3.1?", the agent automatically searches for relevant information from the internet and ultimately generates the following image:

For more advanced users, you can run benchmark evaluations with a one-click script:

After completing the evaluation, you can check the experimental results on the analysis platform:

This article is sourced from: Quantum Bit, original title: "Tencent Open Sources New Framework for Intelligent Agents: No Training Required, No Recharge Needed, Achieving SOTA Agent with Open Source Models"

``

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this is at your own risk