The three planning loops differ on one question: where does the plan live? ReAct holds a plan one step long, replans for free, and gets stuck. Plan-and-Execute commits to a graph up front, which buys parallelism and a cheaper executor and goes stale the moment reality disagrees. Reflexion wraps whole attempts in an outer loop and quietly depends on an oracle you probably do not have. Here is what each costs, with the numbers.
Anthropic measured its orchestrator-worker research system beating a single agent by 90.2%, while burning roughly 15x the tokens of a chat. Cognition spent a year telling people not to build multi-agent systems, then shipped one under a single rule: writes stay single-threaded. Three topologies, the failure mode each one owns, and the one you should default to.
There are thousands of possible coding interview questions and roughly a dozen underlying patterns. Learn to recognize which pattern a question is testing and you stop re-deriving solutions from scratch under time pressure — here is the full list, with what actually signals each one.
The initial design is rarely what separates a hire from a no-hire in an AI system design loop — it is how you handle the follow-up. Here are the five questions that come up constantly across RAG, agent orchestration, and inference serving rounds, with what a strong answer actually sounds like.
RAG pipelines, agent orchestration, and LLM inference serving have become the new bar for senior AI engineering interviews. Here is how to prepare for all three without wasting weeks on the wrong material.