For a long time, I thought technical interviews were broken. If I have public open source code, I thought, that should be enough to prove what I can do. I am not talking about a repository created on Sunday night with three commits called "fix", but about years of projects, issues, pull requests, libraries used by real people and bugs fixed at two in the morning while someone opened an issue containing only "not work".
The ideal interview, in my head, was simple: an engineer would browse my GitHub, pick a project and ask why I had used that structure, why I had rejected another solution and which mistake had forced me to rewrite everything on a Tuesday afternoon. I would explain my choices, they would understand how I think, and then the offer would arrive.
Reality was slightly different. Rapid-fire technical questions, brain teasers, LeetCode quizzes, binary trees to invert on a whiteboard and arrays to sort without consulting the internet. All to prepare you for the real job, where you would spend Monday centering a button and Tuesday in a call deciding whether to name it button-primary or primary-button.
Then the first chatbots arrived in 2022, agents in 2025 and, suddenly, ChatGPT stopped being a tab you kept secretly open and became an app called Codex, permanently open and working on your MacBook. It became an efficient intern: it never takes time off, does not get offended when you ask it to rewrite the same function six times and, most importantly, never says: "The previous colleague who no longer works at the company built this part".
I came across antirez's post, "Control the ideas, not the code". The point is not that expertise no longer matters or that you can type "build me Redis, but prettier" and go to lunch. He describes using AI to implement complex parts of DwarfStar, an LLM inference engine, while focusing on architecture, design, performance, validation and QA instead of reading every function line by line.
He did not stop programming. He moved the work one floor higher.
If a developer can now guide an agent goal by goal, build a huge system and control its quality and direction without having written every single line of code (but above all, and this is what struck me most, without having read it), what should we ask during interviews? "Implement bubble sort in C for me": practically speaking, this is starting to have the same value as asking an architect to manufacture a brick in front of us. To be fair, I would hire them if they managed it.
Meanwhile, Apple has sued OpenAI, accusing it, among other things, of soliciting confidential information while recruiting Apple employees and even asking candidates to bring hardware components to interviews. The allegations have yet to be proven, but the scene is wonderful: for twenty years, they told us to bring a resume. In 2026, apparently, all you need is a prototype logic board in your jacket pocket. If the security alarm goes off, you have probably passed the first stage.
So what should I bring? My prompt history? The conversations in which I guided an agent all the way to a solution? The list of internal processes I automated? Company secrets, perhaps sorted by importance in an Excel file?
Perhaps a 2026 interview should resemble a 2026 job. Give the candidate a small but ambiguous problem and let them use the tools they would actually use, including AI. Watch which questions they ask before starting, how they divide the work, what context they give the agent, how they spot a plausible but wrong answer and what tests they invent to try to destroy it.
Ask why they chose that architecture, what happens when the model hallucinates, how much the solution costs, how they roll back after a bad deployment and which part they would never entrust to AI. Ask them to explain a failure. Successful projects prove that things went well once, but well-explained failures prove that next time might go better. Open source code is still useful, perhaps more than before, but what matters now is the story behind that repository, not the code itself.
In 2026, I do not want to be judged only by how quickly I can sort an array. I want to be judged on my ability to understand whether that array needed sorting, whether we could have avoided creating it, whether the AI agent that sorted it has just deleted the production database and how long it took me to notice.
Our lives as developers need to be rewritten: we have become legacy code that must be retired, and we need to deploy a new version of ourselves.