Anthropic Interview Questions
Curated from verified interview reports: LLM-infrastructure-flavored Coding and OA rounds (tokenizers, in-memory databases, distributed streaming), an Integration round building a Claude tool-calling agent loop, and System Design rounds on LLM inference and chat systems.
18 real interview questions reported from Anthropic Software Engineer interviews (Coding, OA, System Design, Integration rounds): 1 easy, 5 medium, 12 hard. Each question shows its interview round and how often it was reported, and coding questions come with runnable tests.
Anthropic Coding round questions (7)
- Bootloader Loop Detection - Medium, reported 3×, free preview
Simulate a tiny instruction set, report the accumulator before the loop, then flip a single <code>jump</code>/<code>next</code> to make the program halt if possible. - Consecutive Event Detection - Easy, reported 3×, free preview
Find every maximal run of consecutive truthy samples whose length clears a threshold; streaming-friendly logic beats recomputing every window. - Convert Stack Samples to Trace Events - Medium, reported 41×
Turn sampled call stacks into minimal start/end events; the key invariant is comparing adjacent stacks by common prefix, not by set membership. - Distributed Mode and Median - Hard, reported 12×
Combine already-sorted numeric shards to recover global mode and median; the interesting part is tie-breaking and even-length median handling. - File Deduplication by Content - Medium, reported 53×
Group files that share identical content and return only real duplicate clusters; ordering and singleton filtering are the easy-to-miss details. - Longest-Match Tokenizer - Medium, reported 14×
Greedily tokenize text by taking the longest vocabulary match at each position; the trap is correctly handling overlapping tokens and unknown spans. - Multi-Worker Web Crawler - Hard
Crawl only pages under the starting hostname while avoiding duplicate fetches and honoring a concurrency cap.
Anthropic OA round questions (6)
- Banking System with Pending Transfers - Hard, reported 17×
Model accounts, deposits, and two-phase transfers where held funds should not be spendable again until the transfer is accepted. - In-Memory Cloud File Storage - Hard
Build a tiny in-memory object store with add/copy/delete operations plus prefix queries ranked by size, where failure cases matter as much as success cases. - Temporal In-Memory Database - Hard, reported 16×
Implement a record/field store with point-in-time reads, TTL expiration, prefix scans, and deletes that behave correctly across version history. - Recipe Manager with Version History - Medium, reported 6×
Manage recipes with case-insensitive names, ownership checks, and immutable historical versions after each edit. - Task Management with TTL and Snapshots - Hard, reported 5×
Track per-user tasks with expiration and point-in-time snapshot reads; the hard part is making completion and TTL interact correctly across time. - VLIW Kernel Optimization Take-Home - Hard, reported 4×
A rare officially published Anthropic take-home: optimize a tree-walking hash kernel on a simulated VLIW SIMD machine until cycles collapse.
Anthropic System Design round questions (4)
- Design a Batched LLM Inference API - Hard, reported 40×
Anthropic's most-reported design family: serve many LLM requests with dynamic batching, GPU-aware routing, latency SLOs, and graceful behavior under overload. - Design a 1-on-1 Chat System - Hard, reported 7×
A focused realtime-messaging round: presence, online/offline delivery, ordering, and the trade-offs between a lightweight cache bus and a durable log. - Stream a Large File to Many Machines - Hard, reported 2×
Take one inbound copy of a huge artifact and fan it out to thousands of machines without melting the origin link. - System Design Document Review - Hard, reported 3×
Instead of designing from scratch, critique a proposed design doc and surface the missing requirements, failure modes, and operational blind spots.
Anthropic Integration round questions (1)
- Claude Agent Loop with Tool Calls - Hard, reported 3×
Drive a model/tool loop until the assistant stops requesting tools; the twists are multi-tool turns, stable message ordering, and runaway-loop protection.