The Modern AI Stack
Providers, gateways, tools, and platforms — the map of what you assemble versus what you build.
The 2026 stack has settled into layers. Model providers: Anthropic (Claude), OpenAI (GPT), Google (Gemini), plus strong open-weight families (Llama, Qwen, DeepSeek, Mistral) you can host yourself. Above them, gateways and routers (LiteLLM, Vercel AI Gateway, OpenRouter) give you one API over many providers with failover and cost tracking — worth adopting on day one, because you WILL switch models, and hardcoding one provider's SDK everywhere is how a two-line change becomes a two-week migration.
The tools layer is where agents get hands: function calling (last lesson) for your own tools, MCP for standardized third-party ones, and managed capabilities like code execution, web search, and computer use from the providers themselves. The platform layer — LangChain/LangGraph, LlamaIndex, agent SDKs from each lab, plus eval/observability platforms — is optional scaffolding: valuable when your orchestration is genuinely complex, overhead when a 50-line loop would do. Module 3 gives you the decision framework.
Two adjacent practices matter to this course. Agent Skills — folders of instructions and scripts an agent loads on demand — are how you package expertise for coding agents like Claude Code; they're the reusable-knowledge primitive of the agentic toolchain. And vibe coding (Courses 01–03) is how you BUILD all of this faster: the RAG pipelines, eval harnesses, and MCP servers in this course are exactly the kind of software an agent pair-builds well. The tracks compound.