Reducing Batch Makespan for LLM Agent Workloads through Inference-Environment Coordination — ThinkLLM