An agent that starts empty-handed and invents its own tools

What it does

Yunjue Agent tackles open-ended tasks without arriving pre-equipped. It starts from an empty tool library and iteratively writes, tests, and refines its own tools while working. Success or failure of each code execution acts as the only supervision signal, letting the system accumulate reusable capabilities without human-labeled ground truth. The authors back this with open traces and benchmark scripts across five datasets.

The interesting bit

The system treats tool crashes and successes as a self-supervised curriculum, which is a neat way to avoid hallucinated strategy updates. It also batches this evolution in parallel, so the agent is not just solving the task at hand but running a mini population of tool candidates to see which survive. The authors even propose an “Evolutionary Generality Loss” metric to watch convergence, essentially giving evolution a training-loss analog.

Key highlights

• Bootstraps from an empty tool library and reportedly lands second on the HLE leaderboard while outperforming Gemini 3 Pro by 17.4% on DeepSearchQA.
• Uses binary pass/fail signals from code execution to guide tool creation, sidestepping the need for human annotations or static workflows.
• Publishes full system traces, a dedicated reproduction branch, and evolved tool artifacts to make the “black box” auditable.
• Offers a web demo and a CLI skill interface that converts a SKILL.md description into autonomous tool generation.

Caveats

• macOS-only for now; the authors explicitly state it has not been tested on Linux or Windows.
• Requires manual configuration of external services like Codex and Tavily, so out-of-the-box runs are not frictionless.

Verdict

Researchers and agent builders who want to study reproducible, self-improving systems should look here; developers seeking a polished, cross-platform production framework should wait for broader OS support and easier onboarding.