One Anthropic hiring manager says he stopped using traditional coding interviews this month. That is a revealing change in how he wants to assess engineers, but it is not an announcement that Anthropic has abolished coding assessments or permits AI in every interview.
In an October 27 personal essay, Felix Rieseberg argues that conventional exercises have become less useful to him as AI handles more implementation work. His account is explicitly personal, and the replacement assessment is still unfinished. Those boundaries matter as much as the provocative premise.
A manager questions what the exercise measures
Rieseberg says he now gives more weight to judgment, empathy and the ability to build something that addresses a real need. He describes reviewing portfolios and looking for focused prototypes. His preference for initiative over years of experience is advice to applicants, not evidence that seniority has stopped mattering across the engineering profession.
His assessment of Claude’s performance on small coding challenges is also his own observation, not an independent benchmark of candidates against models. Most importantly, he acknowledges that he has not yet worked out the replacement interview process.
That makes the essay an account of a hiring problem rather than a finished hiring standard. If an exercise no longer separates the skills a manager wants, dropping it creates another obligation: deciding what evidence will do a better job. A portfolio can show an outcome, but an interviewer still has to understand the applicant’s contribution and reasoning.
The interesting shift is therefore from producing an answer in a bounded exercise towards explaining choices in a larger piece of work. That is a proposed change in what gets observed, not proof that technical understanding has become dispensable.
Company instructions still define the actual interview
Anthropic’s published candidate guidance, labelled last updated July 10, 2025, is older background and separate from Rieseberg’s new essay. It encourages candidates to use AI for preparation and to refine an application after writing the initial draft themselves. It asks that the experience described remain genuine.
For take-home assessments and live interviews, however, the default instruction is not to use AI unless permission is explicitly given. The guidance says the company will indicate when assistance is allowed. A personal account of changing an interview format does not override the instructions for a particular role.
The company also says that it uses Claude to help with parts of recruitment but does not let the model make hiring decisions. This is its description of its process, rather than an independent audit.
For applicants, these distinctions avoid a false choice between pretending AI does not exist and assuming all evaluation is now AI-assisted. Preparation, portfolio production and a live assessment are different settings. A fair process should make the permitted tools clear before asking someone to demonstrate a skill.
Better demonstrations need better interpretation
There is a separate measurement problem behind this debate. In a February 2026 methodology update, METR explained why its newer developer-productivity experiment had become difficult to interpret. Developers’ willingness to participate without AI and their choice of tasks introduced selection effects; concurrent agents also complicated measurements of working time.
That research is not a hiring study and cannot validate Rieseberg’s approach. It does show why a visible result, or someone’s confidence about working faster, is not automatically a clean measure of capability or contribution.
A prototype-focused interview would face its own version of that problem. The evaluator could ask which alternatives were rejected, what failed during development, and how the candidate checked the final behaviour. These are possible assessment questions, not a replacement process that Rieseberg or Anthropic has announced.
His essay’s immediate significance is narrower and more useful than a sweeping prediction about engineering jobs: one manager has concluded that his old exercise no longer captures what he wants to hire for. Whether a new method identifies those qualities more reliably remains to be demonstrated.










