Yunjue Agent Tech Report: A Fully Reproducible, Zero-Start In-Situ Self-Evolving Agent System for Open-Ended Tasks

TL;DR

The In-Situ Self-Evolving paradigm is proposed, which treats sequential task interactions as a continuous stream of experience, enabling the system to distill short-term execution feedback into long-term, reusable capabilities without access to ground-truth labels.

Abstract

Conventional agent systems often struggle in open-ended environments where task distributions continuously drift and external supervision is scarce. Their reliance on static toolsets or offline training lags behind these dynamics, leaving the system's capability boundaries rigid and unknown. To address this, we propose the In-Situ Self-Evolving paradigm. This approach treats sequential task interactions as a continuous stream of experience, enabling the system to distill short-term execution feedback into long-term, reusable capabilities without access to ground-truth labels. Within this framework, we identify tool evolution as the critical pathway for capability expansion, which provides verifiable, binary feedback signals. Within this framework, we develop Yunjue Agent, a system that iteratively synthesizes, optimizes, and reuses tools to navigate emerging challenges. To optimize evolutionary efficiency, we further introduce a Parallel Batch Evolution strategy. Empirical evaluations across five diverse benchmarks under a zero-start setting demonstrate significant performance gains over proprietary baselines. Additionally, complementary warm-start evaluations confirm that the accumulated general knowledge can be seamlessly transferred to novel domains. Finally, we propose a novel metric to monitor evolution convergence, serving as a function analogous to training loss in conventional optimization. We open-source our codebase, system traces, and evolved tools to facilitate future research in resilient, self-evolving intelligence.

Figures (12)

• Figure 1: An architecture overview of Yunjue Agent.
• Figure 2: Performance comparison of Yunjue Agent against state-of-the-art agents and agentic foundation models. Our method is highlighted in cyan, and the backend model (Gemini 3 Pro) appears in orange. *Gemini 3 Pro denotes our implementation with a Python interpreter.
• Figure 3: Frequency distribution of the toolset evolved across five benchmarks. We report the top 50 tools, illustrating the emergence of high-generalizability primitives.
• Figure 4: Evolution of the tool library size relative to the cumulative number of processed queries. The experimental sequence follows the curriculum HLE $\rightarrow$ DeepSearchQA $\rightarrow$ FinSearchComp $\rightarrow$ xbench-ScienceQA $\rightarrow$ xbench-DeepSearch, highlighting the convergence of tool synthesis.
• Figure 5: Venn diagram visualization of tool correspondence between zero-start and warm-start settings on DeepSearchQA. The left set comprises tools unique to the zero-start baseline ($\mathcal{T}_{\text{DSQA}} \setminus \mathcal{T}_{\text{HLE}}$), while the right set consists of incrementally generated tools in the warm-start setting ($\mathcal{T}_{\text{HLE}\rightarrow \text{DSQA}} \setminus \mathcal{T}_{\text{HLE}}$). Central intersection indicates high functional overlap. Distinct points represent individual tools, arranged by semantic similarity. Tools in the intersection share similar functionalities, while those within the central dashed lines are exact matches.