Felix Rieseberg, an engineering manager at Anthropic who led Claude.ai, Cowork, and desktop Claude Code, has removed traditional programming tasks from his team's hiring process: the model now solves such exercises, not the candidate. Instead of coding challenges, he looks at resumes and portfolios with small prototypes, although he does not yet have a proven replacement for the previous selection mechanism. The practical takeaway for junior engineers is that a working prototype built over a weekend says more about a candidate than proficiency in interview tasks.

image

What happened

In October 2026, Felix Rieseberg, an engineering manager at Anthropic responsible for Claude.ai, Cowork, and desktop Claude Code, stopped giving candidates in his team traditional programming tasks. On October 27, 2026, a post titled 'Hunger beats tenure: Advice for junior engineers' was published on his blog, where he explains the logic behind the decision: exercises like React challenges or questions about assembly were previously a 'weak but useful' predictor, but are now useless because 'Claude is now better at both.' In interviews, Rieseberg looks at resumes and portfolios for small prototypes that show taste, empathy for the user, and the ability to reason from first principles. One example from his team is the graphical interface for Claude Code: Anthony Morris and Alice Zhao, two relatively new engineers, built a working prototype in a single weekend.

Context

Coding tasks have long served as a standard filter in engineering hiring: they did not provide a precise signal, but they replaced some selection criterion where no other verifiable data existed. The problem here is classic for any measurement tool: when the 'solver' outpaces the target population of candidates, the test ceases to distinguish people from one another. Rieseberg calls this event 'another milestone where software engineering reaches a new level of abstraction,' meaning he views it not as a personal whim, but as another step in the devaluation of manual code writing as a measurable skill. It is also important to note that this is the practice of his team, not an announced overhaul of Anthropic's entire hiring procedure.

Why this matters for the industry

For the industry, this is a specific marker of how AI is changing engineering hiring. If a model solves typical interview tasks better than candidates, 'manual coding speed' ceases to be a signal of suitability, and attention shifts to how a person formulates the problem, works with AI agents, and what business result they deliver. ZipRecruiter data confirms this picture: a survey of 1,500 employers in October 2026 showed that 70% of IT employers have already switched to controlled or in-person tests—the industry is not canceling skill verification, but looking for formats resistant to model assistance. A separate product direction is emerging in this space: AI-resistant assessments, proctoring, and prototype evaluation instead of open coding tests. For now, the effect is more of a signal than a mass phenomenon: the abandonment of coding challenges is established in specific teams, primarily within AI companies.

Why this matters for users

Those entering the profession or changing jobs should restructure their preparation priorities. Preparing for 'algorithms and data structures' for whiteboard interviews becomes less important, while showing a working prototype, built even over a weekend, and the ability to find 'friction' in a product is more valuable. Rieseberg advises taking not random tasks from the backlog, but those that move the company's goals forward, choosing one sharp idea, and relying on the fresh perspective of a newcomer. However, it is still too early to completely abandon classic preparation: in most companies, interview formats remain the same, and typical tasks continue to appear in interviews.

What is still unknown / limitations

Rieseberg's post is an expert observation by a practitioner, not a measurement: which tasks, on what data, and by what methodology he compared the results of Claude and candidates is not disclosed. The case of the graphical interface for Claude Code involves two specific engineers without a comparison baseline and without a described contribution from agents, so a single intra-team example cannot be used to derive a stable pattern regarding the compression of the 'idea—product' cycle. The author himself notes that he does not have a validated replacement for the previous selection mechanism, and the spread of this practice beyond his team has not yet been confirmed by independent data.

Sources

Author

Look at AI, editorial team