Why this site exists
Building agents has become a job of its own. Companies now hire engineers specifically to put language models in a loop with tools, memory and real permissions, and they interview for it with questions that general software or machine learning preparation does not cover: when to use an agent at all, how to design a tool for a model, how to keep a long run coherent, how to evaluate a system that behaves differently every time, and how to stop it doing damage when it reads something hostile.
Most material about agents is written to sell a framework or to show a demo. Interviews ask about something else: the mechanisms underneath, the failures in production and the trade-offs between designs. AI Agent Interviews is written for that.
What it covers
The site is organised around twelve areas that make up agent engineering interviews:
- Agent loops and architecture, and when a workflow is the better choice
- Tool design and function calling
- Context engineering
- Agent memory
- Planning and reasoning
- Multi-agent systems
- Agent evaluation
- Safety, security and prompt injection
- Protocols such as MCP and A2A
- Reliability, cost and latency in production
- Coding and computer-use agents
- Agentic system design, including build-an-agent coding rounds
It deliberately stays inside that scope. General machine learning theory, model training and low-level inference serving are covered well elsewhere and are not repeated here.
Editorial standards
- No invented facts. No made-up statistics, no fabricated quotes and no claims that a question was “asked at” a named company unless there is public evidence for it. When public information is thin, the page says so.
- Dated where it moves. Model capabilities, framework behaviour and protocol details change quickly. Anything that depends on a version or a date says which one.
- Mechanisms over products. Answers explain the loop, the state and the guarantee, and mention specific products only as examples.
- Every answer decides something. A good interview answer ends in a choice and its cost. So does every answer here.
- Sources for claims. Where an idea has an origin, such as ReAct, the τ-bench pass^k metric or the term “lethal trifecta”, the page names it.
- Corrections are made in place and the page's updated date changes with them.
What is on the site
- AI agent interview questions: 48 questions across the twelve areas, each with what a strong answer covers.
- The agentic system design interview: a seven-step method, worked end to end.
- AI agent glossary: the vocabulary, defined briefly.
Everything published here is free to read, and no account is needed.