Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence

Abstract: LLMs are increasingly expected to serve as general-purpose agents that interact with external, stateful tool environments. The Model Context Protocol (MCP) and broader agent skills offer a unified interface for connecting agents with scalable real-world services, but training robust agents remains limited by the lack of realistic environments and principled mechanisms for life-long learning. In this paper, we present \textbf{Agent-World}, a self-evolving training arena for advancing general agent intelligence through scalable environments. Agent-World has two main components: (1) Agentic Environment-Task Discovery, which autonomously explores topic-aligned databases and executable tool ecosystems from thousands of real-world environment themes and synthesizes verifiable tasks with controllable difficulty; and (2) Continuous Self-Evolving Agent Training, which combines multi-environment reinforcement learning with a self-evolving agent arena that automatically identifies capability gaps through dynamic task synthesis and drives targeted learning, enabling the co-evolution of agent policies and environments. Across 23 challenging agent benchmarks, Agent-World-8B and 14B consistently outperforms strong proprietary models and environment scaling baselines. Further analyses reveal scaling trends in relation to environment diversity and self-evolution rounds, offering insights for building general agent intelligence.

• The paper presents a novel framework for synthesizing real-world environments and tasks by mining web data and using executable tool validation.
• It employs a closed-loop RL system with diagnosis-driven self-evolution, achieving significant improvements on diverse tool-use benchmarks.
• Empirical results demonstrate strong cross-environment generalization and robust scaling laws with performance gains over 20 percentage points.

Agent-World: Real-World Environment Scaling for General Agent Intelligence

Introduction

"Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence" (2604.18292) systematically addresses the twofold challenge of scaling diverse, realistic tool-use environments for language agent training while simultaneously enabling continuous, closed-loop self-improvement via diagnosis-driven environment-task evolution. The framework advances beyond static or purely LLM-driven simulated settings, establishing an agent training and evaluation infrastructure rooted in real web-mined data, executable tool ecosystems, and dynamic RL-driven curriculum.

Figure 1: Overview of Agent-World (left) and its downstream agent performance scaling (right) across key benchmarks MCP-Mark, BFCL V4, and $\tau^2$-Bench.

Real-World Agentic Environment Synthesis

Agent-World’s environment synthesis pipeline employs three pivotal advances: large-scale autonomous theme extraction, web-scale structured data mining, and tool interface generation with robust verification. Starting from thousands of domains (sourced from real-world MCP server specs, open-source tool documentations, and PRDs), the system’s deep-research agent leverages retrieval, browser, code, and OS tools to extract and organize domain-aligned databases encompassing heterogeneous file types (json, csv, sql, etc.).

Tools are synthesized and unit-tested in-the-loop by code generation agents, with rigorous compilation and execution-based filtering. Only tools achieving executable, validated behavior (accuracy threshold $>0.5$) are retained. Each synthesized environment, defined as $(\mathcal{D}, \mathcal{F})$ (database and toolset), is systematically categorized into a multi-level taxonomy to enable stratified sampling and analytic control for downstream training and evaluation.

Figure 2: Agentic Environment-Task Discovery pipeline, from theme selection to database mining to tool/validator generation and task synthesis.

Figure 3: Hierarchical taxonomy: first-tier (left) and second-tier (right) distribution of synthesized environments in Agent-World.

Figure 4: Statistical overview of the Agent-World ecosystem, detailing the scale, diversity, and difficulty of environments and tasks.

Dynamic, Verifiable Task Synthesis

Built atop the real-world environment layer is a task synthesis module supporting two complementary paradigms: graph-based (sequential tool dependency) and programmatic (complex, possibly non-linear workflows). Tool graphs encode strong/weak/independent dependencies between tools, with random walks producing chain-of-action traces serving as the ground truth for task generation. Tasks are formulated without exposing tool specifics and are paired with structured rubrics or executable verification code, all subject to sandbox evaluation and stability checks. Difficulty is explicitly controlled through increased sequence complexity and obfuscation, with pass@10 rates confirming substantial challenge even for proprietary LLMs.

Continuous Self-Evolving Agent Training

Agent-World’s core training methodology is a closed-loop, multi-environment RL system based on the Group Relative Policy Optimization (GRPO) algorithm. The agent interacts with an evolving curriculum of environments and tasks, receiving rewards from executable rubrics or validators that enforce not only answer correctness but also state consistency and procedural compliance.

A central innovation is the integration of an explicit diagnosis agent into the RL loop. After each round of training and dynamic evaluation (in stratified arenas), the diagnosis agent localizes failure modes by analyzing execution traces, error statistics, and environment metadata. Weaknesses trigger targeted environment or task complexification (database enrichment, harder queries) in the next RL cycle, forming an automated, self-improving curriculum (Figure 5 and Figure 6). This co-evolution of agent and environment reflects curriculum learning at ecosystem scale, validated quantitatively by round-over-round monotonic improvements.

Figure 5: Schematic of continuous self-evolving RL training: rollout, open-loop evaluation and diagnosis, then targeted environment-task expansion.

Empirical Results and Scaling Analysis

Agent-World, instantiated in 8B and 14B scale models, was extensively benchmarked against advanced proprietary (e.g., GPT-5.2 High, Claude Sonnet-4.5, Gemini-3 Pro, Seed2.0) and open-source models (e.g., Qwen3-235B, DeepSeek-V3.2-685B, EnvScaler, AWM) over 23 agentic tool use, advanced assistant, software engineering, research, and general reasoning benchmarks.

Salient findings:

• Superior cross-environment generalization: Agent-World-8B/14B consistently surpassed open-source environment scaling baselines and outperformed larger foundational LLMs on realistic tool-use evaluations such as MCP-Mark, BFCL V4, and $\tau^2$-Bench.
• Positive, nontrivial scaling laws: Increasing training environment diversity from 10 to $\sim$2,000 led to substantial downstream performance gains (>20 percentage points), with progress sustained, albeit with diminishing marginal returns, at higher scales.

  Figure 7: Downstream agent scores scale positively as the number of synthesized training environments increases.

• Continual self-evolution effectiveness: Each self-evolution cycle yielded monotonic accuracy improvements on all core agentic tool-use benchmarks, especially in settings requiring state tracking and multi-step orchestration.
• Robustness on long-horizon, heterogeneous domains: Agent-World-8B achieved leading radar plot profiles on general reasoning, agentic search/coding, and knowledge/MCP domains, acquiring transferable agentic strategies without benchmark-specific tuning.

  Figure 8: Consistent generalization for Agent-World-8B across general reasoning, agentic search/coding, and knowledge/MCP benchmarks.

• Advanced assistant transfer: Agent-World series demonstrated superior transfer on SkillsBench, ARC-AGI-2, and Claw-Eval, outperforming both parameter-scaled LLMs and environment scaling methods.

  Figure 9: Comparative generalization of Agent-World and baseline models on advanced AI assistant evaluation suites.

• Stable RL dynamics: Training curves for both 8B and 14B backbones displayed consistent reward growth and controlled entropy, indicating robust, non-collapsing exploration in diverse environment space.

  Figure 10: RL training dynamics—reward and entropy trajectories—of Agent-World-8B/14B under multi-environment RL.

Implications and Future Directions

Agent-World establishes a blueprint for scalable agentic RL in realistic, compositional, and stateful tool-use environments, with continuous self-evolution bridging the gap between static curriculum and open-ended, co-evolving intelligence. This paradigm is directly extensible to multimodal, broader API, or real-world embodied settings and supports the empirical investigation of scaling laws under environment-driven curriculum learning.

Several implications arise:

• Agent-environment co-evolution as curriculum: Systematic, diagnosis-guided environment complexification is critical for robust transfer and compositional generalization, directly addressing previous limitations of one-pass, static agent RL.
• Executable task-grounding as a regularizer: Reliable, verifiable reward signals serve as a practical alternative to noisy or hallucinated feedback in simulation-only pipelines, mitigating reward hacking and distributional shortcuts.
• Scalable annotation and automatic validation: The synergy of LLM-based generation, rigorous code verification, and evaluator agents provides a scalable path to constructing and evaluating realistic agentic benchmarks and RL training corpora.

Future research may extend the framework to:

• multimodal (vision-grounded, tactile) agentic tasks,
• life-long open-ended self-improvement,
• scalable, robust alignment via continual environment/diagnosis integration,
• and unified benchmarking across real industry toolchains and synthetic sandboxes.

Conclusion

Agent-World (2604.18292) advances the frontier of LLM agent training by providing a framework that jointly scales realistic environment-task synthesis and implements diagnosis-driven, self-evolving RL. By tightly integrating web-mined, executable environments with dynamic task, reward, and curriculum construction, and by demonstrating competitive generalization and robust scaling profiles across challenging domains, Agent-World offers a principled infrastructure for the next generation of general-purpose intelligent agents.