A couple months ago, I went on the CTO Confessions podcast. In one of the segments, we talked about the impending—or already developing—problem of the shrinking talent pipeline. With agentic AI, senior engineers are now expected to manage multiple coding agents to augment their output, and this is in lieu of more traditional training of junior engineers recruited from coding schools or computer science programs. This isn't hypothetical; the data from the last 5–6 years shows a significant reduction in hiring junior engineers, to the point where students are responding by choosing other majors.
The Skill Code: How to Save Human Ability in an Age of Intelligent Machines addresses this concern directly. The author had spent years researching how humans learned on—or as it turns out, sometimes off—the job. But as AI continued to show extraordinary progress, he reframed his findings to highlight the importance of preserving skills across generations of humans, the transference of experience between novices and experts.
The book breaks it down into three determinants, and names them: Challenge, Complexity, and Connection. In my own career as an engineering leader developing people on my teams, these categories certainly resonate:
Challenge is setting up tasks so that a novice can learn by doing something unfamiliar, and ideally, unsettling. People on my teams will know this as "Growth via Discomfort," where I'll intentionally assign work so they'll be forced to do something they've never done before and even be scared of trying. This could be running an all-hands; pushing code to production; or writing up a post-mortem. When the task is tough, but not so hard as to be overwhelming, the novice makes incremental improvement.
Complexity is providing space to the novice so that they can discover the solution for themselves. This is reminiscent of the pedagogical approach taken up by educators when I worked in EdTech—provide the framework, but also give students some autonomy to figure things out for themselves. This model helps the learner independently arrive at their own solutions, with a sense of achievement that makes the lessons more memorable.
Connection describes the social bond between mentor and mentee. Having the teacher around allows them to give feedback and encouragement, and this earned familiarity is what allows mentors to craft the right challenges, at the right times, with the appropriate amounts of complexity. I suspect that this aspect of skill development is the hardest for AI to replicate; so far all of our AI usage shows that humans strongly prefer interacting with other humans.
This explicit breakdown shows that skill development is multifaceted, but also interdependent: it's the design of tasks, their incremental difficulty and ambiguity, but also in an environment conducive to learning. We just lived through this in the pandemic aftermath. Remote work was initially prevalent and unavoidable, but it was the junior employees who wanted to go into the office for more face time, so they could socialize and build those connective bonds with their more senior colleagues.
Experts also suffer when mentorship ties are severed: first by COVID, then by AI. They lose the intrinsic satisfaction of mentorship, but also the opportunity to improve by teaching others. The act of mentoring others is not just a form of altruism and paying it forward; the mentor also gets to practice their skills, adjust and refine their approach based on mentee feedback and outcomes. In the best cases, they elevate their mentees to take over aspects of their responsibilities, which frees up space for the mentor to take on new challenges themselves.
Breaking this with AI sacrifices all of these benefits. Worse, it threatens to widen the skill gap. The book flags this problem—massive skill inequality—where the experienced workers execute with prompts and specs, while the less experienced workers don't receive the support needed to self-improve. The demand for software engineers has continued to tick up, but positions are now tilted towards more experienced roles. It's an echo of the K-shaped economy that economists identified during the pandemic recovery, featuring a stark bifurcation of the haves and have-nots.
The Skill Code identifies what it takes to transfer skills from one generation to the next: challenge, complexity, and connection. The emergence of AI disrupts this framework by looking to replace aspects of this skill code, but that's a choice borne of short-term efficiency, not a long-term inevitability. AI in education provides an early blueprint—instead of just replacing people with automation, we can leverage AI as a tutor, using its flexibility to fill in gaps in current employee training systems that are often stale and too rigid. If we reframe our goals to be the rich inheritance of skill across generations, then AI's role would naturally shift to be an augmentative accelerant.
