ATLAS: Automated TCO-Oriented LLM Migration with Agentic Systems

Rapid AI accelerator development drives Large Language Model (LLM) owners to migrate models to new hardware to optimize Total Cost of Ownership (TCO). However, this migration is complex, involving model analysis, profiling, compilation and run-time validation---a multi-step workflow often error-prone, requiring weeks of senior engineer effort. This paper introduces ATLAS, an agentic system enabling TCO- and performance-driven workload mobility. ATLAS is a multi-agent framework that leverages LLM agents to analyze model architectures, recommend optimal hardware configurations, and propose migration solutions. Additionally, ATLAS integrates Model Context Protocol (MCP) tools to directly interact with hardware systems, knowledge bases, and hardware simulators, achieving a complete, end-to-end migration system. By replacing manual migration and performance optimization with an agentic workflow, ATLAS significantly accelerates the migration lifecycle. Evaluations demonstrate that this framework improves workload TCO by up to 50.8\% and throughput by up to 64.84\%, while reliably preventing compilation and runtime failures before physical deployment.