Frameworks Landscape: LangChain, LlamaIndex, CrewAI, AutoGen, Semantic Kernel, Assistants API

Interviewers rarely ask "which framework is best" — that question doesn't have a stable answer and everyone knows it. What they actually ask, in one form or another, is: what would this framework hide from you, and when would that hurt? That is a system-design question dressed up as a tooling question, and it rewards the same kind of reasoning as the rest of an AI-engineering interview — trade-offs, failure modes, and "what would you do instead" — not a feature-by-feature comparison or a recitation of which package has more GitHub stars this quarter.

This subject treats every framework as an answer to a narrower question: what does it abstract away, and what does that abstraction cost you when something goes wrong in production? A framework that hides retry logic, context assembly, and the exact tokens sent to the model is exactly the kind of "magic" you are being interviewed on — because the interviewer wants to know you can still reason about latency, cost, and failure modes underneath it, not that you can name the class that wraps a chat completion call. Framework APIs also change every few months; a memorized method signature ages out fast, but the underlying trade-off (composability vs opacity, control vs velocity, self-hosted vs managed) does not.

The goal here is not to teach you to use any of these frameworks — that belongs in a tutorial, not an interview. It is to give you a durable mental model for the landscape: a small number of distinct abstraction categories (general-purpose orchestration, RAG/data frameworks, role-based multi-agent, conversational multi-agent, enterprise SDKs, hosted managed platforms), what problem each was built to solve, and the criteria a senior engineer actually uses to decide between a framework and a hand-rolled orchestration layer. It builds on agent-architectures-and-the-agentic-loop, rag-architecture-end-to-end, multi-agent-orchestration, and mcp-and-tool-integration-protocols — read those first if you haven't; this subject assumes you already know what a retrieval pipeline, an agentic loop, and a multi-agent handoff look like, and asks where the frameworks in this landscape try to package them.

By the end of this subject you should be able to place any framework you've never heard of — including whatever launches next quarter — into one of the categories below within thirty seconds of hearing what it does, and reason about its trade-offs without having used it. That transferability is the actual skill being tested. A candidate who can only discuss the six named frameworks has memorized a list; a candidate who can categorize a seventh, unfamiliar one on the spot has understood the landscape.