For a decade, "put it on Lambda" was one decision. It's now four - Functions, Managed Instances, MicroVMs, Durable Functions - with different cost curves, different latency profiles and different limits, all sitting behind the same mental model in most people's heads.

One agentic app spans all of them: bursty embeddings, a streaming gateway idling on CPU, isolated code execution, and orchestration waiting on a human. This session maps shape to tier - where the curves cross, what you give up, and when AgentCore or plain Functions wins instead.