The landscape of building AI agents has shifted dramatically by 2026. The old playbook of assembling frameworks, RAG pipelines, and glue code has largely been replaced by SDKs (Claude Agent SDK, OpenAI Codex SDK) that ship with built-in tools like file read/write, bash, grep, and web search. Three key changes drove this: built-in tools eliminated the need to reimplement common tooling, the skills pattern replaced upfront tool registries with progressive disclosure, and longer context windows demoted RAG from the default retrieval strategy. Frameworks like LangGraph, Pydantic AI, and CrewAI still have their place for multi-provider routing, multi-agent orchestration, deterministic typing, and production observability — but the recommendation is to start with the SDK and only reach for a framework when you genuinely outgrow it. The post concludes with a pitch for Pulumi Neo as an SDK-first agent for infrastructure-as-code use cases.
Table of contents
The old playbookThe shift: three things changed at onceWhen you still need a frameworkWhere this lands for infrastructure workStart with the SDKSort: