Engineering design optimization processes often operate through a collaborative cycle between designers and computational tools, where iterative adjustments implicitly embed human-preference criteria into the final output. However, deliberately shaping the design space remains challenging for an engineer because it depends on expert-defined constraints and parameters, whose original assumptions become increasingly obscured as they pass through multistage pipelines—from initial topology optimization (TO) setup to downstream manufacturability, performance, and aesthetic post-processing [1-2]. To address this challenge, our work explores an architected multiagent AI framework in which multiple domain-specialized Large Language Model (LLM) agents are orchestrated to coordinate and advance toward a unified design objective. We instantiate vision–language agents with ability to interact with the TO landscape by invoking TO solvers and auxiliary design tools, translating user-specified formulations into executable TO code implementations and refining them through experimental adjustment of initial parameters based on user feedback. We evaluate long-horizon design by issuing abstract natural language prompts that require sustained, multi-step reasoning, such as requests to increase or decrease structural complexity by tuning internal branching of the structure. Results are assessed through a combination of quantiative analysis through 3D medial-axis skeletonization and qualitative review through a AI-VLLM judge. The AI judge evaluated 70% of attempts to simplify structures as successful, while the remaining 30% executed successfully but did not achieve measurable simplification. In all simplification cases, AI judgements agreed with human baseline evaluations and were consistent with skeletonization-based complexity metrics. In contrast, agents struggled with performance on complexity-increasing tasks, with only 40% judged successful. Among the failures, 67% produced non-runnable TO code, and among the successes, half showed disagreement between AI judgment and skeletonization metrics. All remaining AI judgements were consistent with skeletonization results. Overall, our framework demonstrates the capabilities and limitations for multi-agents to translate long-horizon user intent into topology optimization workflows.