What top companies want (2026)

A Java job posting says "5 years Spring, strong SQL" and mostly means it. An "AI Engineer" posting in 2026 is noisier — RAG, agents, evals and inference optimization all compete for the same title, and the loop that tests them differs by company. This page is the sourced signal underneath the noise: two studies of real postings, and how five kinds of companies actually run their loops.

sourced 2026-09-27 ~10 min 2 studies, 4,600+ postings 5 loops
⚠️ Hiring facts change fast. Every figure below is dated and sourced. Treat it as a starting point, not a guarantee — check the loop and the bar with your own recruiter before an interview.

What top companies actually ask for

One analysis of 1,000+ AI-engineer job descriptions found 95.6% production-oriented roles — most "AI engineer" postings want someone who ships, not someone who trains a model from scratch. Python appeared in 82.5% of postings, AWS in 40%, model training in only 6.4% (AI Shipping Labs). A separate study of 3,647 postings at 198 AI companies — OpenAI, Anthropic, Databricks, Waymo among them — found a sharper picture of which skills matter most (aijobprep):

Share of 198 AI companies asking for each skill

👉 Every bar below is a real number from aijobprep's count, not an estimate. Tap a skill for what it means in a posting and why it ranks where it does.
Python & LLMs
83%
Agents
81%
Observability
78%
Evals
68%
Distributed sys
62%
Inference opt.
57%
RAG
40%
Tap a skill above.
Takeaways. Python and LLM fundamentals are close to universal (83%); agents, observability and evals now outrank RAG on its own. A portfolio that only demonstrates RAG is optimizing for the smallest of these seven bars.

How the big loops actually test you (2026)

The skills above tell you what to practice; the loops below tell you how each company checks for it (Wrok, Perspective, Sundeep Teki).

Five loops, one tap each

👉 Tap a company or loop type; predict its distinguishing round before you check.

Meta

AI-enabled coding round.

Google

Hiring-committee packet review.

Amazon

Leadership Principles, every round.

Anthropic Applied AI

Five stages, customer discovery.

Forward-deployed (OpenAI/Anthropic/Palantir)

Scoping under constraints.

Tap a loop above.
Expert note. The GenAI system-design round (IGotAnOffer) follows a five-step framework — framing → architecture → deep dive → trade-offs → wrap-up — and interviewers probe evals and regressions, token budgets and caching, prompt-injection defence, chunking and embeddings, and rate limits and fallbacks. AI Architect and Solutions Architect roles (AWS, Google Cloud, Microsoft) add enterprise requirements, cross-cloud reference architectures and build-vs-buy, FinOps cost modelling, governance and compliance, and presenting trade-offs to non-engineer stakeholders. The 12 worked designs and a timed-practice mode for this framework are AJ3, planned next.

Check yourself