Towards General Agentic Intelligence via Environment Scaling (2509.13311v1)

Abstract: Advanced agentic intelligence is a prerequisite for deploying LLMs in practical, real-world applications. Diverse real-world APIs demand precise, robust function-calling intelligence, which needs agents to develop these capabilities through interaction in varied environments. The breadth of function-calling competence is closely tied to the diversity of environments in which agents are trained. In this work, we scale up environments as a step towards advancing general agentic intelligence. This gives rise to two central challenges: (i) how to scale environments in a principled manner, and (ii) how to effectively train agentic capabilities from experiences derived through interactions with these environments. To address these, we design a scalable framework that automatically constructs heterogeneous environments that are fully simulated, systematically broadening the space of function-calling scenarios. We further adapt a two-phase agent fine-tuning strategy: first endowing agents with fundamental agentic capabilities, then specializing them for domain-specific contexts. Extensive experiments on agentic benchmarks, tau-bench, tau2-Bench, and ACEBench, demonstrate that our trained model, AgentScaler, significantly enhances the function-calling capability of models.

• The paper presents a novel method for constructing diverse, verifiable simulated environments that enhance function-calling capabilities in language models.
• It implements a two-stage agent experience learning protocol that achieves state-of-the-art performance even in compact model regimes like AgentScaler-4B.
• Empirical results on multiple benchmarks validate the framework’s effectiveness, while also revealing challenges in long-horizon tool calling and domain specialization.

Advancing General Agentic Intelligence via Environment Scaling

Introduction

The paper "Towards General Agentic Intelligence via Environment Scaling" (AgentScaler) presents a systematic framework for developing general agentic intelligence in LLMs by programmatically scaling and diversifying simulated environments. The central thesis is that robust function-calling capabilities in agents are tightly coupled to the diversity and verifiability of the environments in which they are trained. The authors introduce a pipeline that automates environment construction, synthesizes agentic tasks, and implements a two-stage agent experience learning protocol. The approach is validated on multiple agentic benchmarks, demonstrating strong performance, particularly in compact model regimes.

Environment Construction and Scaling

The environment scaling methodology is grounded in the abstraction of function calls as read–write operations over domain-specific databases. The pipeline begins with large-scale scenario collection, aggregating over 30,000 APIs from public and internal sources. These APIs are filtered and refined to ensure explicit input–output specifications and compositional compatibility.

A tool dependency graph is constructed, where nodes represent tools and edges encode parameter-based compositional relationships. Louvain community detection is employed to partition the tool graph into coherent domains, each associated with a database schema that formalizes the environment's state space. Tools within each domain are programmatically materialized as executable code, enabling direct manipulation of the underlying database.

Figure 1: Overview of the automatic environment build and agentic task construction pipeline, illustrating scenario collection, tool graph modeling, and programmatic materialization.

Agentic task construction proceeds via forward simulated agent–human interplay. Initial environment states are sampled for diversity, and tool sequences are generated by traversing the domain-specific tool graph. Each step involves argument generation and tool invocation, with state transitions tracked for verifiability at both the database and tool-sequence levels.

Agent Experience Learning

Agent experience is collected through simulated human–agent interactions within the constructed environments. A simulated user with a predefined intent interacts with the agent, which must leverage domain-specific tools to fulfill the task. Completed interaction traces are subjected to a rigorous three-stage filtering process: validity control, environment state alignment, and function calling exact match. This ensures high-fidelity supervision and robustness, even retaining trajectories with intermediate tool-call errors if the overall intent is achieved.

Figure 2: The agent interacts with the simulated user and changes the environment state through generated functions, enabling scalable experience collection.

The agentic experience learning protocol is two-phased. The first phase focuses on general tool-usage competence across broad domains, while the second phase specializes the agent in vertical domains with domain-specific tasks and tools. The training objective masks out human instructions and tool responses from the loss, propagating gradients only through assistant-generated tool calls and responses, thereby conditioning the model on context while optimizing for agentic actions.

Experimental Results

AgentScaler models (4B, 8B, 30B-A3B) are trained on Qwen-3 backbones and evaluated on $\tau$-bench, $\tau^2$-Bench, and ACEBench. The results indicate that AgentScaler-30B-A3B achieves state-of-the-art performance among open-source models under 1T parameters, and in several domains approaches or matches closed-source systems. Notably, AgentScaler-4B demonstrates competitive performance with models several times larger, underscoring the efficiency of the environment scaling and experience learning pipeline.

Ablation studies confirm the necessity of the two-stage training protocol: general foundation learning is critical for tool-usage competence, and domain specialization further consolidates these capabilities.

Robustness, Generalization, and Long-Horizon Analysis

AgentScaler models exhibit strong robustness and generalization, as evidenced by out-of-distribution performance on ACEBench-zh. The synthetic data approach enables efficient knowledge transfer, with substantial gains in agentic capabilities for compact models.

Figure 3: Pass$^{\wedge}k$ metric results across all domains in the $\tau^2$-Bench, demonstrating stability and consistency of AgentScaler models.

Stability analysis using the pass$^{\wedge}k$ metric reveals that AgentScaler-30B-A3B consistently outperforms its backbone across all evaluated settings. However, a negative correlation is observed between the number of tool calls in a trajectory and task accuracy, indicating that long-horizon tool calling remains a fundamental challenge for agentic models.

Implications and Future Directions

The AgentScaler framework demonstrates that scalable, verifiable environment construction and principled agent experience learning can yield robust agentic intelligence in LLMs, even at modest parameter scales. This has significant practical implications for deploying agentic models in resource-constrained or latency-sensitive scenarios.

Theoretically, the work highlights the importance of environment diversity and verifiability for generalization and robustness in agentic intelligence. The pipeline's modularity suggests extensibility to broader modalities and real-world deployment.

Future research directions include integrating reinforcement learning atop the simulated environments, scaling to larger model architectures, and addressing the long-horizon tool-calling challenge. The authors also note the potential for edge deployment and broader applicability of compact agentic models.

Conclusion

This paper presents a principled approach to advancing general agentic intelligence via environment scaling and agent experience learning. The automated construction of diverse, verifiable environments and the two-stage training protocol enable efficient synthesis of agentic capabilities, validated by strong empirical results. The framework sets a foundation for scalable, robust, and generalizable agentic LLMs, with clear avenues for further research in RL integration, model scaling, and long-horizon reasoning.