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

TL;DR

Yunjue Agent is a tech report from Yunjue Technology proposing the “In-Situ Self-Evolving” (ISE) paradigm: treat sequential task interactions as a continuous experience stream and let the agent autonomously synthesize, optimize, and reuse Python tools without ground-truth labels, starting from a zero-tool state. The architectural bet is that tool functionality is the right starting point for self-evolution because its feedback signal is intrinsic and binary (execution succeeds or throws), unlike the subjective/delayed rewards of workflow- or context-evolution. A multi-agent system (Manager, Tool Developer, Executor, etc.) plus a “Parallel Batch Evolution” strategy crystallizes runtime feedback into a shared capability pool; the paper claims significant zero-start gains over proprietary baselines (GPT-5, Gemini 3 Pro with web+Python) across five open-ended benchmarks. Companion repo is public; full trajectories promised for 2026-01-31.

Key claims

Among the three pillars of LLM-agent adaptation — workflow, context, tool — the toolset is the foundational prerequisite for novel environments, and tool functionality offers verifiable binary feedback (code runs or throws), making tools the logical starting point for in-situ self-evolution where supervision is scarce [§Introduction].
The In-Situ Self-Evolving (ISE) paradigm treats task interactions as a continuous stream from which the agent distills short-term execution feedback into long-term reusable capabilities, with no access to ground-truth labels and a zero-start initialization [§Introduction, Abstract].
Yunjue Agent is built as a multi-agent architecture (Manager, Tool Developer, Executor, …) that synthesizes bespoke Python primitives at runtime; tools are continuously synthesized, optimized, merged, and reused in a shared capability pool [§Method].
A “Parallel Batch Evolution” strategy is introduced to optimize evolutionary efficiency — multiple tool-synthesis/validation trajectories are explored concurrently rather than sequentially [§Method].
Across five diverse benchmarks under a zero-start setting, Yunjue Agent reports significant performance gains over proprietary baselines (GPT-5, Gemini 3 Pro) which were augmented with web search and Python interpreters; on xSciQA where standard leaderboards prohibit external tools, the backend model is also explicitly equipped with a Python interpreter [§Experiments].
Complementary warm-start evaluations claim the accumulated general knowledge transfers seamlessly to novel domains [§Experiments].
A novel convergence-monitoring metric is proposed to track when the evolution process has stabilized, serving a role analogous to training loss in standard supervised learning [§Method / §Experiments].

Method

The system instantiates the ISE paradigm as a small constellation of role-played agents on top of an LLM backbone. A Manager decomposes the incoming open-ended task; a Tool Developer writes Python primitives on demand; an Executor runs them in a sandbox and feeds back success/exception signals. Tools that succeed get crystallized into a shared capability pool with metadata, similar tools get merged, and failing tools get either repaired or discarded — all driven by the binary execution signal rather than by external labels. The starting condition is “zero-start”: the pool is empty, and the agent must synthesize every tool it eventually uses.

Parallel Batch Evolution runs multiple synthesize→validate→merge rounds concurrently rather than serially, exploring several tool variants per task to accelerate convergence. The paper proposes a convergence metric that monitors when the capability pool stops growing/changing, functioning as a stopping criterion analogous to training-loss plateaus in supervised learning.

Evaluation is on five open-ended benchmarks under a zero-start setting; baselines include GPT-5 and Gemini 3 Pro augmented with web search + a Python interpreter. On xSciQA — where standard leaderboards prohibit external tools — Yunjue’s backend model is explicitly equipped with a Python interpreter for a like-for-like comparison.

Results

The paper claims (per the public abstract and front matter):

Significant gains over proprietary baselines (GPT-5, Gemini 3 Pro with web + Python) across the five-benchmark suite under zero-start.
Complementary warm-start transfer experiments where the accumulated tool pool ports cleanly to novel domains.
A reported convergence signature from the proposed metric demonstrating stability of the evolved pool.

Specific benchmark numbers are stated in the report but were not parseable from the abstract/intro fetch; refer to the PDF (https://arxiv.org/pdf/2601.18226) for the headline tables.

Why it’s interesting

This is an explicit bet that tool evolution — not workflow or context optimization — is where open-ended self-improvement should be rooted, justified by the verifiability asymmetry of executable code. That framing dovetails with the closed-loop trajectory-synthesis recipe in Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing (which also exploits binary tool-execution feedback as the supervision signal) and with the in-context “experience library” approach in Training-Free Group Relative Policy Optimization (which similarly avoids weight updates and lets the agent accumulate knowledge across tasks). Yunjue’s distinctive moves are (a) the zero-start initialization — the agent doesn’t begin with a curated toolset at all — and (b) the Parallel Batch Evolution strategy, which is a different axis from MiroThinker’s sequential interaction-depth scaling (MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling) and from K2.5 Agent Swarm’s parallel sub-agent orchestration. The companion repo plus the promised release of full trajectories on 2026-01-31 makes this one of the few cleanly reproducible self-evolving-agent reports.