Skip to content
Tech Interview Prep

Real Interview Questions for the Agentic AI Engineer at Amazon

New

Overview

How the loop runs

Candidates describe a funnel, not a single day. A recruiter screen opens it: one reported sample question is "What hands-on experience do you have with post-training or agents?": followed by an online assessment that, since 2025, has included an AI-assisted debugging task. After that comes a set of roughly one-hour rounds.

The only round-by-round account in the reports is for a Software Engineer on an agentic AI security team in Austin (October 2026): three one-hour rounds. A DAG topological sort, a second directed-graph problem, and a third the write-up truncates before naming, though the same account bills the set as graphs, system design and Leadership Principles. Treat that as directional for this title rather than identical to it; the title differs.

Other reports put the agentic content in a technical screen framed as a conventional system design question, an agent running detection over about 1 TB of security telemetry per day, and describe loops spanning Amazon AGI, AWS Applied AI, Annapurna Labs, Alexa and Advertising. A Bar Raiser round appears in the process descriptions, with questions about a major technical decision and its rejected alternatives.

Where reports are thin: coding. One compilation explicitly declines to assert a coding requirement for Amazon AI roles because its retained sources did not clear its evidence bar, while other reports list from-scratch Adam, TF-IDF search, sliding windows and topological sort. Plan for coding as possible, not guaranteed.

What interviewers are evaluating

Not a tool list. The judgement is whether you can bound an agent (cost, latency, blast radius) and then show it works. Nearly every design question in the reports arrives with a constraint attached: 200 ms p99 at 10,000 QPS, TTFT under 200 ms for a 70B model, 20% of the originally allocated compute, 1 TB/day of telemetry. What is being bought is your ability to turn that number into an architecture and defend the arithmetic behind it.

Three focus areas

Agent architecture under a stated constraint. Tool use, working memory and long-term storage over unstructured domain data; multi-agent orchestration with state persistence and loop termination; adapting mid-flight when requirements shift without resetting system state. The scored part is what you cut when compute drops to a fifth, and which failure you choose to accept.

Evaluation and failure diagnosis. Online and offline pipelines that benchmark agent output and feed corrections back; self-correcting loops built on user corrections, tool outcomes and execution traces; debugging an agent that invents API parameters during tool calling. Symptom, hypothesis, instrument, fix, not "add guardrails".

Model and retrieval depth. Scaled dot-product attention with a causal mask plus tests that would expose future-token leakage; custom PyTorch blocks; inference bottlenecks and serving tradeoffs; HNSW versus IVF-PQ at billions of vectors; vector versus keyword retrieval. Depth here is checked in code and complexity, not vocabulary.

Behavioral rounds run against the Leadership Principles alongside the technical ones: a failed architectural decision and its reversal, tradeoffs under severe resource limits, a time AI gave you a wrong answer.

How to prepare

Answer aloud against a clock: state the constraint, do the capacity arithmetic, name the failure mode, then the design. Work this card's Interview Q&A in order: design questions run through most of it, with behavioral at 18–20 and coding at 24–25, and use the Practice MCQ to check the mechanics you would otherwise hand-wave: attention masking, index tradeoffs, loop termination. The role's own prep covers the concepts behind each.

Feedback on this overview

Curated: · Written: · Reviewed: