Sierra AI Just Reinvented the Software Engineering Interview — And It’s Brilliant
Sierra AI — the enterprise AI startup co-founded by Bret Taylor (OpenAI board chairman and former Co-CEO of Salesforce) — just published a quietly revolutionary blog post titled The AI-native interview. In it, they explain how they completely overhauled their hiring process for engineers in the age of powerful coding agents.
The change wasn’t driven by hype. It was driven by reality: traditional LeetCode-style interviews have been broken for a while, and AI tools (Codex, Claude Code, etc.) have made them even more irrelevant.
The Old Way (Still Used Almost Everywhere)
- 2 coding interviews;
- 1 algorithms / ML interview;
- 1 system design interview;
- 1 culture / behavioral interview.
It tested syntax, algorithms, frameworks, and “culture fit.” It was standardized, easy to calibrate, and… increasingly disconnected from what engineers actually do day-to-day in 2026.
The New AI-Native Process
Sierra collapsed the first three technical interviews into one single onsite session with three phases:
1. Plan
The candidate and interviewers sit together and define a product to build. The candidate drives the ideation process; interviewers ask probing questions to push the thinking deeper. The goal is to see real product thinking in the candidate’s domain of expertise.
The interviewer leaves the room. The candidate gets two hours to actually build the product using whatever AI tools, frameworks, and approaches they want. They are explicitly encouraged to pivot, cut scope, or change direction if they hit a wall.
3. Review
The interviewer returns. The candidate presents the working product, walks through key decisions, explains the architecture (data model, abstractions, extensibility), and discusses the path to production and scaling.
Before the main onsite, they replaced the old “remote coding phone screen” with a new debugging / feature-addition task: candidates are given a medium-sized real codebase + a draft PR from a colleague and must load it, understand it, run it, and improve the feature.
Result: the first stage tests “how do you work inside an existing codebase?”, while the main session tests “how do you create something meaningful from scratch?”
Why This Is a Much Better Signal
- More representative of real engineering work today;
- Higher signal — it reveals initiative, ownership, judgment under time pressure, and product sensibility;
- Way more engaging — many candidates told them it was the most fun interview they’ve ever had.
They openly admit the format is harder to standardize, so they counter that with:
- Clear, outcome-agnostic evaluation rubrics;
- Interviews run in pairs for better calibration;
- Pre-sharing evaluation criteria and tips (e.g., “it’s totally fine to cut scope”).
They even tested it with infrastructure engineers and concluded it works there too.
Also read:
- OpenAI's Shadow Library Scandal: Deleted Datasets and the High Stakes of AI Copyright Evasion
- xAI’s Efficiency Crisis: 11% MFU, a Founder Exodus, and a $60 Billion Hail Mary on Cursor
- LLMs Explained: How Large Language Models Work (and Where They Commonly Fail)
The Bottom Line
In an era where agents can already solve most LeetCode problems in minutes, asking someone to write a binary tree inversion by hand tells you almost nothing useful. Asking them to conceive, build, and defend a real product in two hours using modern tools tells you everything.
Sierra’s new process feels like the first genuinely AI-native hiring system I’ve seen from a serious company. It’s not perfect, but it’s honest, thoughtful, and clearly designed around how great engineers actually work now.
If more companies follow suit, we might finally move past the era of artificial whiteboard interviews and into something that actually predicts who will thrive.
What do you think — would you want to go through this kind of interview as a candidate? I’d love to try it.
Full post (highly recommended): The AI-native interview — Sierra