Abstract : The modern AI agent has outgrown the language model that gave rise to it. Where earlier systems were measured by their ability to predict the next token in a sequence, today's agents reason through multi-step problems, invoke external tools, browse live web environments, execute and debug code, and refine their plans in response to real-world feedback. These capabilities do not come from any single technique; they emerge from a coherent pipeline of interdependent training stages that must be co-designed to work effectively together. This paper surveys the AI agent training stack through five distinct layers: (1) foundation pretraining, which establishes the base model's knowledge and in-context learning capacity; (2) instruction and preference alignment, which shapes behavioral compliance through reinforcement learning from human feedback (RLHF), Constitutional AI (CAI), and direct preference optimization (DPO); (3) reasoning-oriented post-training, which amplifies structured problem-solving through chain-of-thought (CoT) prompting and reinforcement learning from verifiable rewards (RLVR), achieving results such as 79.8% on the AIME 2024 mathematics competition; (4) tool use and environment interaction, evaluated on benchmarks including SWE-bench, WebArena, and tau-bench; and (5) distributed systems infrastructure, encompassing tensor and pipeline parallelism, ZeRO optimizer sharding, and IO-aware attention kernels. For each layer, the paper examines current methods, evaluation frameworks, infrastructure constraints, and open challenges. The central finding is that agent training is a co-design problem: architecture, data, inference-time computation, and infrastructure must be optimized as an integrated system rather than in isolation.