Yunjue Agent is the official in-situ self-evolving agent system from Yunjue Technology, designed to convert runtime execution feedback into reusable tools for open-ended tasks.
Open assets: GitHub · Traces Dataset
We follow an in-situ self-evolving paradigm: each execution outcome becomes a supervision signal, and tools are continuously synthesized, optimized, merged, and reused in a shared capability pool.
Sequential tasks are treated as a live experience stream. No ground-truth labels are required for capability accumulation.
Tools serve as verifiable primitives with binary execution feedback, reducing strategy drift and improving reliability.
Aggregator and Merger components cluster and consolidate tools, improving evolution efficiency at scale.
Evolutionary Generality Loss (EGL) is used to monitor stabilization, analogous to training loss in optimization.
The following figures and values are summarized from the public tech report and project assets provided by Yunjue Technology.
Interpretation: the system demonstrates both initial capability bootstrapping and cross-domain reuse through evolved tools.
| Dataset | Reported Yunjue Agent Score | Setup | Source |
|---|---|---|---|
| HLE | 48.0% | Zero-start | Tech Report (arXiv:2601.18226v2) |
| DeepSearchQA | 73.5% | Zero-start | Tech Report (arXiv:2601.18226v2) |
| FinSearchComp | 65.0% | Zero-start | Tech Report (arXiv:2601.18226v2) |
| xbench-ScienceQA / xbench-DeepSearch | Improved vs. backend baselines | Zero-start + transfer | Tech Report + project README |
The system uses coordinated roles for planning, tool generation, execution, and integration, with an absorbing mechanism to update the global tool pool.
These demos illustrate tool decomposition, PDF retrieval, stock information tasks, and skill-driven automation workflows.
Example: decomposing a user request and generating task-specific tools during runtime.
Example: reusing evolved tools to answer US stock-related queries.
Example: converting structured skill descriptions into executable tool workflows.
Example: locating local files, summarizing documents, and producing presentation artifacts.
To align with E-E-A-T principles, we provide transparent authorship, dated updates, public artifacts, and clear source references for key claims.
Tech report authors include Haotian Li, Shijun Yang, Weizhen Qi, Silei Zhao, Rui Hua, Mingzhu Song, Xiaojian Yang, and Chao Peng, affiliated with Yunjue Technology.
Correspondence: qiweizhen@yunjuetech.com
Short answers for common implementation and evaluation questions.
Yunjue Agent evolves tools during inference and keeps successful tools for future tasks, rather than relying only on a fixed toolset defined before deployment.
Zero-start means the system begins with an empty tool library and builds capability through runtime tool synthesis, validation, and reuse.
Use the public repository, reproduce branch, and released trace datasets to inspect execution logs and rerun benchmark workflows.
Tool evolution is the primary capability engine: tools are synthesized, executed, optimized, and absorbed into a shared pool when they prove useful.
Parallel Batch Evolution processes multiple tasks in batches and uses aggregation and merging steps to reduce redundant tools and improve evolution efficiency.
Evolutionary Generality Loss (EGL) is used as a convergence signal to track whether the evolved tool library is becoming more stable and reusable.
Yes. Warm-start evaluations reported in the project materials indicate that previously evolved tools can transfer to new domains with reduced marginal tool growth.
The project provides both a local web demo and a CLI skill demo. The published quick-start notes indicate the current tested environment is macOS.
The repository is released under the Apache License 2.0, enabling transparent use and reproducibility under standard open-source terms.