Anthropic Infrastructure Software Engineer Interview Experience
Anthropic · Staff · Software Engineer
Anthropic was very different from the 20 companies I interviewed with. The culture round felt like a lawyer call, very interrogative, with conscience based questions like when have you done something against your values, and after that round I was like, is it still worth it?
Interview process
I interviewed for an infrastructure software engineer role that was basically staff-scoped even though the title stayed generic. The process was a recruiter screen, a live coding phone screen, then two onsite loops: system design, coding, culture, and later experiences/goals plus a technical project deep dive. The phone screen was the only question I had seen before online. Everything else felt genuinely novel and very tied to GPU infrastructure and Anthropic's safety culture. The culture interview in particular was unusually interrogative, and they even warned me in advance that some questions might make me uncomfortable. I felt strongest in the project deep dive and weakest in the later coding round.
Interview rounds · 4
- 1
Recruiter screen
BehavioralI had many reschedules for this round. The recruiter explained the staff-level scope, told me references would only happen near offer stage, and asked me to choose between two compensation bands before the loop.
Q1. Which compensation band works better for you?
How they answeredThey framed it as two comp options for the same infrastructure SWE role and said they'd figure out placement from the interview. In hindsight, it was a pretty unusual tactic because I was basically anchoring myself before I even had an offer.
- 2
Phone screen
CodingData Structures & AlgorithmsTechnicalThe phone screen was a live coding round in Replit. It was the only Anthropic question I had actually seen before online, so I came in expecting it. The base problem was straightforward, and then I used the remaining time to push it toward multithreading.
Q1. Build a web crawler that crawls a website.
How they answeredI wrote a basic crawler from a seed page against their small toy website in Replit and walked the links recursively. I normalized the URLs by stripping extra query-string detail, stored the unique links in a dictionary, and printed that out. I got to a working baseline pretty fast, then started talking through and implementing a multithreaded version. We ran out of time there, but the interviewer basically said he thought I could finish it.
Follow-up questions- Strip out extra URL details like query-string noise and print only the unique links.
- How would you make it multithreaded?
- 3
Final / onsite round
System DesignCodingBehavioralArtificial IntelligenceMy first onsite loop had a system design round, another coding round, and a culture screen. The technical interviews felt very gotcha-driven in a practical way, like they already knew the hard edge cases from real infra work and wanted to see if I'd surface them myself. The culture round was the most distinctive part of the whole process because it felt unusually interrogative about mission, safety, and conscience.
Q1. Design an end-to-end system that batches lots of single-line user queries onto GPUs that can process up to 100 queries at once.
How they answeredI proposed queueing requests and flushing either when I hit a batch-size threshold or a time threshold, then dequeueing that batch and sending it to a GPU. The harder follow-up was GPU selection. He kept drilling into how I'd know which GPU had capacity, so I suggested a GPU-aware load-balancing layer that tracks availability and routes work accordingly. It felt like a genuinely novel problem, and even he seemed to be thinking through the tradeoffs with me.
Follow-up questions- When do you decide to flush a batch?
- How do you figure out which GPU has capacity and which one should receive the next batch?
Q2. Given profiler samples with function start and end events, figure out which part of the code is the slowest.
How they answeredThe input was a stream of profiler samples with function start and end markers, and I had to infer enough structure to identify the slowest section. I had practiced a similar stack-sample problem before, but this version kept shifting. Once I had a solution path, he started breaking it with edge cases like cycles, infinite loops, and functions that never return. I think this was my weakest round because I wasn't proactively naming all those failure modes the way they'd expect at staff level.
Follow-up questions- How would you handle cycles or an infinite loop?
- How would you account for a long-running function that never returns?
Q3. Why Anthropic, and what did you take away from the safety and AI security material we sent you?
How they answeredI came in having read the safety and AI security posts they had sent, so I talked through my interpretation of their mission and why they care so much about safe AI standards. After that it became a conscience-heavy discussion about values, pushback, and mission versus profitability. It honestly felt less like a normal behavioral and more like being grilled by legal. They had even warned me ahead of time that some questions might make me uncomfortable.
Follow-up questions- Do you really believe in the mission?
- Would you act on your values or optimize for profitability?
- Can you push back if something goes against your values?
- Have you ever done something that felt against your values?
- 4
Final / onsite round
BehavioralProject DiscussionCross-FunctionalExecutionSystem DesignMy second onsite loop was an experiences and goals interview with an engineering manager and then a two-interviewer technical project deep dive. The first part felt more familiar, closer to Amazon-style behavioral signal, while the deep dive was much more about defending real technical decisions at staff depth. I felt better about this loop than about the coding round from loop one.
Q1. What kinds of experiences, teams, and projects have been good or bad fits for you, and what made you want to switch?
How they answeredI walked through the kinds of teams and projects where I do my best work, the situations that made me want to move on, and what I think my sweet spot is. A lot of it felt like Amazon-style behavioral questions in disguise: conflict, collaboration, goal-setting, KPIs, and whether I had to change the success criteria midstream with cross-functional partners. It was less surprising than the culture round, but still pretty probing.
Follow-up questions- What project did you like most and least?
- How do you collaborate with a colleague when there's conflict?
- Did you meet the original project goals, or did you have to adjust the goals or KPIs during execution?
- How did you define success criteria with cross-functional partners?
Q2. Present a project you led end to end that shipped.
How they answeredThey pushed on the exact things you'd expect at deeper system-design depth: multi-region rollout, sharding, and operational reliability. That round felt good for me because I had actually built the system and could answer the follow-ups concretely.
Follow-up questions- Walk us through the high-level design and why you chose decision A over decision B.
- How would you take it to multiple regions globally?
- How would you shard it?
- How would you make it operationally sustainable and reliable?
Tips from the candidate
I'd prep in two lanes. First, do the common Anthropic phone-screen coding question because that one really is out there. Second, don't expect the rest to come from a bank. I'd read their safety and AI security writing closely, have my own real opinions on it, and for staff-level coding and design I'd make sure I'm the one volunteering the edge cases before the interviewer has to feed them to me. I'd also expect novel infra questions around batching, GPUs, and operational tradeoffs, not textbook system design.
Company culture
I got the sense they're hiring for people who can handle novel infra problems and who genuinely buy into the safety mission. Compared with the 20 or so other companies I interviewed with, Anthropic was the most interrogative by far, especially on values and conscience. They seemed to care a lot about whether I'd push back if something felt wrong, not just whether I could execute. The process also felt pretty structured: I got handed from one recruiter to a more experienced one after the phone screen, references were only supposed to happen near offer stage, and they even talked comp bands before final leveling.