Agent toolkits are expanding fast: prompts, system prompts, agent skills, MCP servers, sub-agents, and declarative workflows all compete for a place in your architecture. Each has genuine strengths, and API practitioners increasingly need a clear sense of when to reach for which. There is also a second, quieter problem: agents change behavior even when you don’t change them. Models get updated underneath you, and yesterday’s reliable prompt is today’s regression. This talk presents a reliability ladder — from prompts and system prompts, through agent skills, MCP tools, and specialized sub-agents, up to declarative Arazzo workflows — as a joint framework for two questions API practitioners face every day: which construct is the right home for this piece of logic, and how do I keep agents maintainable when the models beneath them shift? Each rung trades LLM discretion for reproducibility, and each moves maintenance pain into a different place. You will leave with a decision framework for choosing the right construct per task, a sharper view of where model drift actually hurts, and practical guidance for structuring agents so they keep working when tomorrow’s model arrives.