Large-scale LLM agent serving suffers from resource fragmentation due to task-level resource allocation. We present Wenbo, a scalable agent reasoning system based on dynamic DAG scheduling. Wenbo consists of three components: a Director that incrementally constructs execution DAGs, a Scheduler that globally dispatches ready nodes based on resource requirements, and heterogeneous Workers that execute nodes on GPUs and CPUs. Our theoretical analysis shows that the potential speedup depends on the balance between resource availability and workload demands. For example, with one GPU and 1,024× CPUs, the potential speedup can be 1,024 for CPU-heavy workloads with suitable DAG structures. In addition, we introduce ARGB, an agent reasoning graph benchmark comprising real-world workloads and configurable synthetic workloads. Experiments show that Wenbo achieves 1.4–3.8× higher throughput than Flash-Searcher on real-world workloads, with even much larger gains on synthetic workloads. Our code is available at https://github.com/wenbo-project/wenbo.