What if your AI agent started with ZERO tools… …and had to build its own capabilities to solve problems? No preloaded tool library. No fixed set of capabilities. Just an agent that can create tools, test them, improve them, and reuse what works. Meet Yunjue-Agent — an open-source research system exploring self-evolving AI agents. Here's how it works: → Start from zero: The agent begins with an empty tool library. → Break down complex tasks: It decomposes problems into smaller operations it can execute. → Build its own tools: It generates tools to perform tasks such as searching and scraping PDFs. → Verify through execution: Successful and failed tool runs provide feedback for improvement. → Reuse what it learns: Useful tools become reusable capabilities for future tasks. The goal isn't just to give an AI more memory. It's to let the agent expand what it can actually do. And the researchers didn't just publish a demo. They released code, benchmark scripts, evolved tool artifacts, and full interaction traces so others can inspect and reproduce the experiments. Their evaluations cover five benchmarks, including HLE, DeepSearchQA, FinSearchComp, and xbench. They also report a 17.4 percentage-point gain on DeepSearchQA over Gemini 3 Pro under their zero-start evaluation setup. That's a project-reported benchmark result, not a guarantee of general superiority. The big question this research raises: What happens when an AI agent doesn't just learn how to use tools — but learns how to create the tools it needs? REPOO👇