AgentSkiller: Scaling Generalist Agent Intelligence through Semantically Integrated Cross-Domain Data Synthesis

TL;DR

AgentSkiller introduces a scalable, DAG-driven framework to synthesize high-fidelity, cross-domain, long-horizon interaction data for training generalist LLM agents. The pipeline builds domain ontologies, entity graphs, service blueprints, policies, and cross-domain task fusion, then generates and validates executable task environments with automated rollouts. Empirical results from roughly $\approx 11{,}000$ trajectories show that models trained on AgentSkiller data—particularly the AgentSkiller-14B variant—achieve strong performance on function-calling benchmarks, rivaling many proprietary models and outperforming open baselines, especially in cross-domain, multi-turn settings. This work demonstrates that semantically integrated data synthesis, guided by a deterministic state-machine and rigorous validation, can substantially boost generalist agent capabilities and scalability in real-world, multi-domain contexts.

Abstract

Large Language Model agents demonstrate potential in solving real-world problems via tools, yet generalist intelligence is bottlenecked by scarce high-quality, long-horizon data. Existing methods collect privacy-constrained API logs or generate scripted interactions lacking diversity, which struggle to produce data requisite for scaling capabilities. We propose AgentSkiller, a fully automated framework synthesizing multi-turn interaction data across realistic, semantically linked domains. It employs a DAG-based architecture with explicit state transitions to ensure determinism and recoverability. The pipeline builds a domain ontology and Person-Centric Entity Graph, defines tool interfaces via Service Blueprints for Model Context Protocol servers, and populates environments with consistent databases and strict Domain Policies. A cross-domain fusion mechanism links services to simulate complex tasks. Finally, the pipeline creates user tasks by verifying solution paths, filtering via execution-based validation, and generating queries using a Persona-based Simulator for automated rollout. This produces reliable environments with clear state changes. To demonstrate effectiveness, we synthesized $\approx$ 11K interaction samples; experimental results indicate that models trained on this dataset achieve significant improvements on function calling over baselines, particularly in larger parameter regimes.

Figures (9)

• Figure 1: Performance on $\tau$-bench, and $\tau^2$-bench and ACEBench-Agent. Specifically, our AgentSkiller-14B achieves significant improvement on $\tau^2$-bench and ACEBench-Agent.
• Figure 2: The End-to-End Single-Domain Synthesis Framework. The pipeline progresses through Ontology Definition (Steps 1--5), Executable Implementation (Steps 6--8), and Task Instantiation (Steps 9--17). Crucially, the rightmost column details the Execution-Based Task Filtering and Multi-Dimensional Evaluation, ensuring only valid, high-fidelity environments are retained.
• Figure 3: The Semantic Cross-Domain Fusion Mechanism. Building upon Figure \ref{['fig:single_domain']}, this phase (Steps 11--13) synthesizes composite scenarios. The left panel illustrates Trajectory Fusion via shared entities, while the right panel depicts Policy Harmonization to merge conflicting rules into a unified governance policy.
• Figure 4: Data Theme Analysis. (a) The dataset prioritizes complex operational topics like Logistics and Healthcare. (b) The framework generates a dense web of cross-domain connections (center), resulting in a Cross-Domain subset (right) that is significantly larger than the Single-Domain subset (left), extending the interaction horizon beyond isolated tasks.
• Figure 5: Complexity Analysis. Cross-domain trajectories demonstrate significantly higher complexity in turn depth and tool usage compared to single-domain tasks.