Career in the age of AI agents: what really counts, per Phil Chen

2026-07-05

In the age of AI agents, the most valuable work becomes the kind you cannot grade with a loss function. That, in short, is the thesis of Phil Chen’s essay “Career advice in the age of AI”, written by the founder and former engineer at Helm AI, Scale AI, OpenAI and Google DeepMind. Chen notes that AI models keep getting better at anything you can frame as a well-defined problem with a checkable answer, which is exactly what school teaches. The valuable work of the coming decade, he argues, is everything that cannot be checked during a model’s training. I read the piece through the lens of our industry and picked out what matters most for people early in their careers, and at the end I add my own comment on what it means for recruitment.

Why school stops being enough

Chen writes as the founder of a fully agent-native company, one where no one writes code by hand, and his hiring needs differ from anything he knew before. Over six years he worked in organizations of every size, from his own startup, through Helm AI (which grew from 15 to 50 people), Scale AI (from 500 to 1,500), OpenAI (from 1,500 to 3,000), up to Google with more than 100,000 employees. His conclusion is that since school is mostly well-defined tasks graded against known answers, those are exactly the skills an agent takes over first, and the edge comes from what the machine cannot quantify.

Truly scarce resources: time, relationships and reputation

Chen’s first piece of advice is about where to direct your attention. He recalls that before joining Scale he had offers with clearly higher guaranteed cash, but chose Scale for the people and the exposure to different products. It was that network of contacts, not the extra money, that later opened the doors to DeepMind and OpenAI and gave him a community of founders. His claim is that capital is easier to access today than ever, while what stays truly scarce is time, strong relationships and a documented track record of good work. So Chen advises doing good work and making sure other respected people know about it, and ruthlessly ordering your priorities so your time goes to problems you consider meaningful. He adds that in the age of vibe-coding it is easy to find ways to make a quick buck, but the bigger reward usually waits where you look for real value.

The skill that gains the most: finding problems

The thread of the essay that hits me hardest is that what will count is not only solving problems, but above all finding them. Chen describes how his company searched for signal among candidates and concluded that, since no one writes code by hand, classic tasks that test coding ability, or even system design questions, no longer correlate with real work. Instead the team built interviews that check how quickly someone understands the environment they land in, how they identify problems worth solving, and how they execute within existing constraints. He predicts that the most important skills will be those tied to problem selection and resource allocation, because as ever more capable agents handle complex, well-defined tasks, the greatest impact will belong to the people best at pointing to the important problems and allocating tokens and time to them. Chen also notes that students can feel discouraged that an agent solves all their problem sets, yet in interviews there are huge differences in how much time and how many tokens a candidate needs to reach a solution, and the best ones bring intuition and outside context to their work with the agent.

Ambitious problems and the last mile

Chen’s next pieces of advice fold into a single idea: aim high and polish to the end. He recalls the “bitter lesson” from AI research, which says that scaling general methods ultimately beats task-specific optimizations, and applies it to choosing problems and companies. Since building software is now easy and almost anyone can assemble simple systems, durable value emerges only from extreme focus on genuinely ambitious problems. Hence his advice to judge whether a company is working on the most ambitious version of its problem and whether it has a real shot at solving it. The second pillar is the last mile. Chen cites Alfred Lin’s piece that the final 10 percent is at once 90 percent of the work and 90 percent of the reward, because the median result today is whatever an agent produces from a sloppy prompt, so value comes from a unique perspective and attention to detail.

Scoring goals, not just seeing chances

Chen illustrates this with his own decisions. In 2023 he turned down offers from Anthropic (around 50 people at the time) and Cursor (then two non-founder employees), because he wanted to work on model training and inference at DeepMind, and in 2024 he turned them down again to join OpenAI. Each of these would have been high xG, that is, a high expected goal from a career standpoint, but he bet on fit with his own interests and culture. Chen also does not believe that superintelligence will replace all knowledge work, because humans have an edge in choosing meaningful problems and in allocating capital to solve them. When assessing early-stage companies he advises looking at the team and the market rather than the current product, which almost always changes a lot, and he recalls that Anthropic’s first demo was a Slack bot worse than ChatGPT.

Being a researcher is a mentality, not a job

The last thread is about breaking into research. Chen argues that today you can start without working at a top AI lab, by using the models and turning your own intuitions into evaluations, and that many compute providers give credits to academics. In his view being a researcher is a mentality, not a job, because a researcher’s work is mostly curiosity to explore new ideas, wrestling with infrastructure, understanding the whole system well enough to debug efficiently, and being able to justify the value of the results. Most ideas will fail at scale anyway, and understanding those failures is the first step toward knowing what actually works.

What this means for recruitment and assessing candidates

Here I step in from my own perspective, as CEO of Element. If Chen is right, and in large part I think he is, then technical recruitment built on algorithmic puzzles loses its point, while what gains value is judging how a candidate moves through an unfamiliar environment, how they pick problems worth attention, and how they collaborate with agents. For HR teams and recruiters this means shifting from the question of whether someone can solve a given task to whether they can find the right task and carry it to the finish. That is harder to measure, which makes an organized process matter even more, one where you can see a candidate’s work history, their projects and how they think, not just the score of a single test. This resonates with me all the more because, ever since vibe-coding let us speed up the development of Element considerably, I mostly think about which further problems our ATS should solve, rather than how to do it or how long it will take.

The bottom line

Chen’s essay ends on an optimistic note: the world is still full of opportunity, and the key is to seek interesting problems and deliver extraordinary results. I agree, and I will only add that for recruiters the same challenge applies to the way they assess people. If you are building a team in a world of agents, it is worth asking yourself now whether your recruitment process measures the finding of problems, or still only their solving.