Coinbase has completely redesigned its engineering hiring process: classic rounds of coding from scratch have been replaced by an assessment of AI-Fluency — a candidate's ability to direct AI, verify its output, and apply expert judgment where models make mistakes. By the fourth quarter of 2025, more than half of the code in the company is generated by artificial intelligence, making the skill of working with AI assistants a selection criterion on par with architectural thinking.
What Happened
Coinbase launched a new technical interview process consisting of three signals: practical work with a real company repository and debugging, system design using AI assistants, and an assessment of leadership skills. The overhaul was phased — a pilot in frontend teams in the second half of 2025, expansion to backend in January 2026, and a company-wide launch in March 2026. Simultaneously, the publication recorded a metric: the share of AI-generated code at Coinbase grew from 5.7% in the first quarter of 2025 to over 50% in the fourth quarter, while humans continue to review approximately 100% of all changes.
Context
The share of AI code has grown almost ninefold in one year, indicating the practical maturity of code generation models — Claude, GPT-4.1, Gemini, Cursor — for mass integration into the development workflow. However, 100% human review means that the quality of generated code has not yet reached a level where automatic validation can be trusted: models still introduce logical errors that require human detection. The three-phase restructuring strategy reflects technical limitations: frontend code is easier for AI due to fewer dependencies and more predictable patterns, while backend requires an understanding of distributed systems, where AI generation errors are more costly.
Why This Matters for the Industry
The Coinbase post is the first detailed public report from a major public technology company about a complete overhaul of its hiring process for AI code generation. This sets a reference point for the industry: Canva, Meta, and Datadog are already testing similar approaches, while Google has returned to in-person rounds to verify skills. Coinbase's publication accelerates the trend — it is expected that interview preparation platforms (CodeSignal, HackerRank) will begin launching AI-Fluency assessment modules, and a new product category — AI-Fluency Assessment Platforms — may replace classic algorithmic testing platforms.
Why This Matters for Users
The strategy of preparing for interviews by memorizing LeetCode is losing effectiveness. The key skills now being assessed are: prompt quality, the ability to find errors in AI code, understanding model limitations, and architectural judgment. The three-signal AI-Fluency framework published by Coinbase provides specific criteria that can be used for self-assessment — engineers can evaluate their level in three areas: practical work on real repositories, system design with AI, and the ability to make decisions under uncertainty.
What Is Still Unknown / Limitations
The metric of AI code growth from 5.7% to over 50% does not separate volume and quality: with 100% human review, a significant portion of the generated code likely requires manual correction. It is unknown how widely the new process has already been applied in real hiring and what results it shows compared to the old one. Data on other companies (Canva, Meta, Datadog) is provided without detail — the extent and nature of their AI-Fluency experiments are not disclosed.
Sources
- Interviewing Engineers in the AI Era: Lessons from a Year of Rebuilding — Coinbase Engineering Blog
- Tech Interviews in the AI Era: The Rules Are Being Rewritten — Eason Tech Talk
Author
Look at AI, editorial team