Yunjue-Agent

Project Positioning

Yunjue Agent is the intelligent agent system implementation officially open-sourced by Yunjue Technology, with the goal of building self-evolving artificial general intelligence (AGI) and wearable devices. The team believes that true intelligence lies not only in how much past knowledge is stored, but also in the ability to adapt, learn, and create tools when facing an unknown future.

Core Concept: The In-Situ Self-Evolving Paradigm

Traditional agent systems often perform poorly in open-ended environments where task distributions continuously drift and external supervision is scarce, because they rely on static tool sets or offline training, resulting in rigid and unclear capability boundaries. Yunjue Agent proposes the "In-Situ Self-Evolving" paradigm, treating continuous task interactions as a continuous stream of experience, enabling the system to distill short-term execution feedback into long-term reusable capabilities without requiring real labels. Within this framework, the team identifies tool evolution as the key path to capability expansion, because it can provide verifiable binary feedback signals. Based on this, the system iteratively synthesizes, optimizes, and reuses tools to address new challenges, and introduces the "Parallel Batch Evolution" strategy to optimize evolution efficiency.

Main Features

1. Zero-Start SOTA Performance: Starting from an empty tool library (Zero-Start), relying only on generation, verification, and induction at inference time, it achieves leading results on multiple benchmarks. According to the materials, it improves by 17.4% over Gemini 3 Pro on DeepSearchQA and ranks 2nd on the HLE leaderboard, proving the feasibility of bootstrapping general capabilities from zero.
2. Tool-First Evolution Principle: Prioritize tool evolution rather than memory or workflows as the primary driver of capability. Tools provide objective binary feedback through code execution success or failure, serving as a reliable internal supervision signal when manual annotation is lacking, helping to mitigate hallucination risks, prevent policy deviation, and ensure the stable accumulation of general primitives.
3. Fully Reproducible and Open Trajectories: Provide end-to-end code, benchmark scripts, versioned tool artifacts, and complete interaction trajectories, transforming "black-box" agent results into transparent, auditable research, facilitating analysis of tool convergence, evolution efficiency, and merging strategies.
4. Transformation from Skills to Actions: By providing only a SKILL.md (high-level experience is still human-driven), the agent can autonomously generate the tools needed to execute that skill, achieving the transformation from documented knowledge to runnable automation. Example skills are located in the repository's example/cli/skills directory.
5. Evolution Convergence Metric: Propose a new metric for monitoring evolution convergence, functionally analogous to training loss in traditional optimization.

Benchmark Performance

The system was evaluated on benchmarks such as HLE, DeepSearchQA, FinSearchComp (T2 and T3), xbench-ScienceQA, and xbench-DeepSearch, and achieved SOTA results. In addition, supplementary warm-start evaluations confirm that the accumulated general knowledge can seamlessly transfer to new domains.

Environment and Configuration

• Prerequisites: Python 3.12 or higher; package manager uv; operating system MacOS.
• Configuration Instructions: docs/configuration_reference.md explains the meanings of key fields in .env (such as TAVILY_API_KEY, MAX_WORKER_RECURSION_LIMIT, MAX_TASK_EXECUTION_CNT, PROXY_URL) and key fields in conf.yaml (such as VISION_MODEL, SUMMARIZE_MODEL). Configuration templates can be started from .env.example and conf.yaml.example.

Reproduction and Demonstration

• Reproduction: The reproduction code has been migrated to the stable branch reproduce; please refer to that branch for detailed reproduction guidelines. System trajectories are available on Hugging Face for analysis.
• Web Demo: Developers can deploy it themselves to demonstrate Yunjue Agent's tool self-evolution capabilities and execution process. The UI address is http://127.0.0.1:8000/，健康检查地址为 http://127.0.0.1:8000/health，详细指南见 docs/web_demo.md.
• CLI Skill Demo: simply provide SKILL.md, and the agent can autonomously generate tool execution skills. For a detailed guide, see docs/cli.md, and for example skills, see example/cli/skills.

Update Log

• 2026-01-26: First release, open-sourcing the Yunjue Agent system.
• 2026-01-31: Data release, releasing system logs for five benchmark datasets under a zero-start setting (Google Drive), and adding a Hugging Face dataset that can be analyzed with one line of code; also organized the evaluation scripts and reproduction workflow.
• 2026-02-08: Technical report update, adding theoretical and experimental analysis on system performance, cost, and evolutionary generality loss (EGL), accessible via arXiv or the local PDF.
• 2026-02-11: Reproduction migrated to a dedicated stable branch reproduce; released two demos (Web Demo and CLI Skill Demo).

Applicable Scenarios and Limitations

This system is suitable for scenarios that require agents to autonomously expand their capabilities in open-ended environments where the task distribution continually changes and manual annotation is lacking, and is especially suitable for researching self-evolving agents, tool generation and reuse, and reproducible evaluation. In terms of limitations: it is currently only tested on MacOS, and other operating systems may encounter issues; running it requires Python 3.12+ and uv; some features depend on external API keys (such as TAVILY_API_KEY). The project is licensed under the Apache License 2.0.