Towards General Agentic Intelligence via Environment Scaling

TL;DR

This work tackles the scarcity of agentic trajectories by proposing principled environment scaling to create fully simulated, diverse tool-use environments and a two-stage agent experience learning workflow. The authors automate environment construction via a tool graph, domain partitions, and domain-specific databases, while grounding tasks in verifiable read–write operations. A two-phase fine-tuning regime first builds general tool-usage skills and then specializes to vertical domains, achieving state-of-the-art results among open-source models under 1T parameters on τ-bench, τ²‑Bench, and ACEBench, with notable robustness and generalization. The study also highlights ongoing challenges in long-horizon tool calling and discusses future directions including RL on simulated environments and scaling to larger models.

Abstract

Advanced agentic intelligence is a prerequisite for deploying Large Language Models 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.

Figures (5)

• Figure 1: The overview of the environment automatic build, and agentic task construction.
• Figure 2: The agent interacts with the simulated user and changes the environment state through the generated functions.
• Figure 3: Performance comparison on the Normal, Agent, and Overall subsets of ACEBench-en for two-stage training models.
• Figure 4: Pass$\textasciicircum$ k metric results across all domains in the $\tau^2$-Bench.
• Figure 5: Accuracy by tool call count on $\tau$-bench.