Core Insights - The article discusses a new paradigm in AI agents that can autonomously create tools to fulfill tasks without human intervention, showcasing significant advancements in self-evolving capabilities [1][2][21]. Group 1: Agent Capabilities - The new agent, powered by Gemini 3 Pro, demonstrated superior performance in the Humanity's Last Exam (HLE), achieving scores nearly 20 points higher than other agents using disclosed methods [2][12]. - This agent can generate tools on-the-fly, creating 128 unique tools during its evaluation across various benchmarks, indicating a self-sufficient evolution process [12][13]. - The agent's performance improved significantly with the use of previously created tools, demonstrating a clear trend of diminishing returns after reaching a stable number of tools [13][15]. Group 2: Evolution Framework - The research introduces a novel framework called In-situ Self-evolving Agent, which allows the agent to evolve during the inference phase without external supervision, relying on internal feedback and past experiences [21][27]. - This approach contrasts with traditional self-evolving methods that depend heavily on pre-defined training and expert supervision, making it more adaptable and efficient in real-world applications [22][24]. Group 3: Tool Utilization - The agent prioritizes tool creation as a means of evolution, which directly influences its capabilities and performance, allowing it to handle a wide range of tasks effectively [36][40]. - The framework emphasizes the importance of tools in determining the agent's operational boundaries, ensuring high-quality feedback through code execution [37][38]. Group 4: Research and Development - The research was conducted by a team from Yunjue Technology, founded by former Alibaba executive Peng Chao, with a focus on wearable general intelligence [53][56]. - The project was completed with a modest budget of 150,000 yuan, highlighting the potential for impactful research with limited resources [60][61]. Group 5: Open Source and Industry Impact - The self-evolving framework is open-source, allowing for community engagement and further development, which could lead to significant advancements in AI capabilities [49][75]. - The article suggests that the integration of this self-evolving agent could address the challenges of cost, safety, and adaptability in AI applications, particularly in consumer-facing scenarios [62][71].

Core Viewpoint - The article discusses a new paradigm in AI agents that can autonomously create tools to fulfill tasks without human intervention, showcasing significant advancements in self-evolving capabilities [1][2][3]. Group 1: Agent Capabilities - The agent can independently evolve and create tools based on task requirements, demonstrating a level of autonomy previously unseen in AI [3][19]. - In a benchmark test known as Humanity's Last Exam (HLE), the agent outperformed others, achieving a score nearly 20 points higher than undisclosed methods that utilized tools [4][5]. - The agent successfully created 128 tools during its evaluation, indicating a robust ability to adapt and generate resources as needed [19][20]. Group 2: Performance Metrics - The agent's performance showed a rapid initial increase in tool creation, stabilizing at 128 tools, which were deemed sufficient for most tasks [28][33]. - A comparative analysis of different strategies revealed that the agent's performance improved significantly with the reuse of existing tools, leading to fewer new tools being created as the task complexity increased [34][35]. Group 3: Self-Evolution Framework - The concept of in-situ self-evolution allows the agent to learn and adapt during the inference phase without external supervision, relying on internal feedback and past experiences [52][53]. - This framework emphasizes the importance of tools as the primary means of evolution, allowing the agent to expand its capabilities dynamically [62][63]. - The agent's architecture includes roles such as Manager, Tool Developer, Executor, and Integrator, facilitating a structured approach to task completion and tool creation [68][71]. Group 4: Industry Implications - The research highlights a shift towards open-source solutions in AI, with the potential for widespread application in various industries, particularly in scenarios requiring adaptability and low operational costs [88][126]. - The findings suggest that the agent's ability to self-evolve could address challenges in traditional AI models, such as high costs and limited flexibility in handling diverse user needs [106][114].