Anthropic AI Engineer Interview: Rounds, Questions, and How to Prepare

Anthropic does not post a job titled AI Engineer or Generative AI Engineer. The role most people mean when they search for it is Applied AI Engineer, which sits on the Applied AI team alongside Applied AI Architect and Forward Deployed Engineer. Getting the title right matters, because the loop for that role is not the software engineering loop.

The difference is the customer. An Applied AI Engineer builds on Claude with a customer's engineering team rather than building Claude. So the loop keeps the coding and values rounds every Anthropic candidate faces, and adds rounds on LLM implementation and on how you handle a technical conversation with someone outside your own company.

This article covers which role to apply for, what the job posting actually asks for, each round in the loop, the topics worth studying, and a preparation plan. For the standard engineering loop across all teams, read our Anthropic interview process guide.

Which Anthropic Role You Actually Mean

Anthropic organizes engineering into four groups, and the generic phrase "AI engineer" maps onto three different ones. Pick before you prepare, because the loops differ.

If you want to work on prompts, agents, evaluations and production LLM integrations, the Applied AI family is the match. If you want to work on model internals, that is Research Engineer, and it is a different interview with a machine learning bar.

What the Job Posting Actually Asks For

The Applied AI Engineer postings are unusually specific about the skills, which makes them a better study guide than any interview article. They ask for four or more years in a technical or customer-facing engineering role, and the enterprise postings ask for six or more.

The technical requirement is the part worth reading twice. Anthropic asks for production experience with LLMs, naming advanced prompt engineering, agent development and evaluation frameworks, plus strong Python. Not coursework, and not a demo.

The deliverables named in the postings are equally concrete. Applied AI Engineers ship MCP servers, sub-agents and agent skills for production workflows, build evaluation suites and pilots, run technical workshops with customer teams, and feed patterns they notice back to Anthropic's product teams.

Read that list as the syllabus. Every item on it is something an interviewer can ask you to build or critique.

The Interview Loop

Candidates report a loop of roughly seven rounds, run remotely, including for office-based roles.

Reported end-to-end timelines run from about three weeks to about six. The written "Why Anthropic?" field comes before all of it and is read, so treat it as the first round rather than paperwork.

The Technical Screen

Ninety minutes on CodeSignal is the most commonly reported format, with a 60 minute live version for some roles. The problems are not algorithm puzzles. They are multi-tiered build problems: you implement something that works, then the requirements change and you extend it.

The most frequently reported example is implementing a banking system with several transaction types, level by level, where each level must pass its tests before the next unlocks. Candidates consistently report running out of time rather than getting stuck.

Two things follow from that format. Write the simple version first and make it pass. Then keep your structure loose enough that the next requirement is an addition rather than a rewrite, because that is what the later levels reward.

Python is the expected language. Know the standard library well enough that you are not looking up method signatures while the timer runs. We cover the assessment format in more detail in does Anthropic use CodeSignal in interviews.

System Design for This Role

The system design round for Applied AI is not the classic "design a URL shortener" exercise. It is LLM application architecture, which is a different skill with different failure modes.

Expect prompts in this shape:

• Design a document assistant over a customer's internal knowledge base, with permissions that differ by user
• Design an agent that resolves support tickets and escalates the cases it should not handle
• Design an evaluation suite that would catch a quality regression before a customer does
• Take an existing customer architecture and say what you would change

The graded skill is knowing which part of the system failed. A wrong answer from a retrieval system can come from bad chunking, a weak re-ranker, a prompt that ignores the retrieved context, or a model that cannot follow the instruction. Interviewers push until you separate those, because that is the daily work.

Three fundamentals carry most of these questions. Retrieval is covered in system design for RAG. Agent architecture is covered in how to design a multi-agent system and what is agentic system design. Since the postings name MCP servers as a deliverable, be able to explain what MCP is and when a tool belongs behind it.

For the serving side, which comes up more in infrastructure loops than applied ones, see designing LLM inference systems. The full treatment of this interview type is Grokking the AI System Design Interview.

The Values Round

This is the round that rejects the most candidates, and it is the one people prepare for least.

It is not a quiz about AI safety, and it is not a test of whether you agree with Anthropic. Interviewers describe a conversation: a time your values were tested at work, a scenario with no clean answer, a trade-off between what a customer wants and what is responsible to ship. The follow-up questions go toward how you felt and what you actually did.

Two failure modes are reported often. The first is uncritical enthusiasm, which reads as not having thought about the risks. The second is a rehearsed answer, which comes out clean and emotionally flat while a real one is messier.

Preparation here is reading and reflection, not scripting. Read Anthropic's published values and the Responsible Scaling Policy, then find the places where you disagree or find the trade-off genuinely hard. Bring one real story where you pushed back on something at work. Our Anthropic behavioral interview questions page covers the story format.

Can You Use Claude While Preparing?

Yes, in most stages, and the rule changed. Anthropic used to ask applicants to write their materials without AI assistance, so older advice on this is wrong.

For an applied role this is worth noticing. The company expects you to work well with Claude on the job, and it still expects you to write code without it under assessment conditions.

How to Prepare

Weeks one and two: close the production gap. Build one small application on the Claude API that uses retrieval, one tool call and a written evaluation set. The evaluation set is the part most candidates skip, and it is named in the job posting.

Week three: agents and MCP. Extend that application into an agent with two or three tools. Write an MCP server for one of them. Be able to explain why you drew the boundary where you did.

Week four: the coding screen. Practice multi-tier build problems in Python under a timer, not algorithm drills. Take a working solution and extend it with a new requirement, three times over.

Week five: design and values. Run through the design prompts above out loud, ending each with how you would evaluate the system. Separately, write your "Why Anthropic?" answer and one honest story about a values trade-off.

Frequently asked questions

Does Anthropic have an AI Engineer job title? No. The closest posted titles are Applied AI Engineer, Applied AI Architect and Forward Deployed Engineer on the Applied AI team, and Research Engineer on the research side. Apply to the one whose work you want, because the interview loops differ.

Is the Applied AI Engineer interview a machine learning interview? No. It tests building production systems on top of models, not training them. You need prompt engineering, agents, retrieval and evaluation, not gradient descent.

What language should I use? Python. It is the expected language in the technical screen and the one named in the job postings.

How long does the process take? Candidates report about three to six weeks once interviews begin. The written application question comes before that and is read carefully.

What fails most candidates? The values round, according to consistent candidate reports. Most people over-prepare the coding round and under-prepare this one.

Do I need customer-facing experience? The postings ask for four or more years in a technical or customer-facing role, and six or more for enterprise postings. Engineering experience with real customer contact counts.

Related reading

• Anthropic interview process: rounds, questions, and how to prepare
• What to expect in the Anthropic system design interview
• How to answer "Why Anthropic?"
• What is the Anthropic interview process like, round by round
• Grokking the AI System Design Interview