Agent Architectures and the Agentic Loop

"Agent" is simultaneously the most over-hyped word in AI engineering and the most reliably interview-tested one. Every vendor slide deck shows an autonomous loop solving open-ended problems; every interview panel has watched a candidate reach for that same loop to solve a problem that a twelve-line script would have handled more cheaply, more reliably, and more testably. The gap between the hype and the interview signal is exactly the point of this subject: the mechanics of the loop itself are a five-minute explanation, but the judgment about when to use one is what separates a senior answer from a junior one.

This subject gives you the loop precisely — observe, decide, act, repeat — and the two named variants you'll be asked to compare, ReAct and plan-act-observe-reflect. It then spends real time on the section interviewers care about most: when an agent is the wrong tool, with the actual arithmetic behind why compounding error rates make long autonomous trajectories fragile. It closes with termination and budget controls and a catalogue of failure modes with concrete mitigations, because "how do you stop it from spiraling" is asked in almost every agent design interview.

Treat this as the conceptual, vendor-neutral companion to two things already on this platform: case-study-coding-agent works through one fully-specified agent end to end (tools, permissions, context management, cost), and plan-and-loop-modes explains how Claude Code specifically exposes autonomy controls (/loop, /goal, plan mode, dynamic workflows). This subject is the layer underneath both — the general shape of the loop and the design principles that apply whether you're building a coding agent, a support agent, or a research agent.