To address the limitations of large-scale language model (LLM)-based agent systems that demonstrate robust performance across a variety of tasks, this paper proposes a Difficulty-Aware Agentic Orchestration (DAAO) framework that dynamically adjusts workflow depth, operator selection, and LLM assignment based on query difficulty. DAAO consists of three interdependent modules: a variational autoencoder (VAE) for difficulty estimation, a modular operator allocator, and an LLM router that considers cost and performance. This enables fine-grained, query-specific inference strategies by leveraging heterogeneous LLMs and dynamically orchestrating workflows. Across six benchmarks, DAAO outperforms existing multi-agent systems in both accuracy and inference efficiency.
