New skills. A new challenge
Three skills have emerged to make this work: delegation, discernment, debugging.
Delegation: knowing what to hand off, and how. Verifiable, well-scoped work; keep enough context to evaluate the result. The more you delegate, the less you learn what you're handing off; understanding that asymmetry is what makes delegation work.
Discernment: taste, judgment, and context become premium because Agents can produce so much. You can't evaluate what you don't understand. Domain expertise and context become increasingly valuable; without depth, Agents expose the gap in the output.
Debugging: root-cause analysis on failed output. Where was the instruction ambiguous? What context was missing? Which inputs were wrong? AI failures usually reveal problems in the horizontal work.
Debugging agent output is usually debugging your thinking, and increasingly your organisation's.
Effective debugging needs the context delegation erodes. Outsource your contextual knowledge and you can't root-cause. You'll know something's wrong; you won't know why.
Each 3D skill carries an erosion risk; together they compound. Junior professionals who delegate before they learn the context become brilliant at getting Agents to produce, but unable to tell when it's wrong. Is the solution deliberate friction? Rotate people through AI-heavy and AI-light work, preserve mentorship on tasks agents could handle, build "show your work" checkpoints that slow throughput but preserve learning.
Are these "3D skills" durable? Will "friction" counteract atrophy? Here's what the research agents have learnt.
-
Chatterji et al., OpenAI with Columbia & Wharton·evidence
Within firms, the least experienced workers route the most work through AI, at the stage judgement is meant to form.
Applying firm fixed effects so that workers are compared with their own colleagues, an administrative-data study finds early-career staff and trainees send eight to nine more weekly messages to enterprise AI than the average user in their firm, while managers and executives send fewer. Because the comparison is within-firm it removes the confound that heavy-adoption firms simply hire differently, the measurement the delegation-before-context claim previously lacked. The population with the least accumulated context is routing the most work through the system at the apprenticeship stage, direct evidence for the atrophy paradox that junior professionals delegate before they learn what they are handing off.
-
Patel et al., BusinessCaseBench (arXiv)·evidence
Frontier models gained 23 points in two years on the expert-graded case reasoning entry-level roles are built on.
Models now clear the expert-written instructor rubrics across the business disciplines, and stumble only on the stricter test of satisfying every criterion an instructor named. The reasoning being cleared, synthesis and judgment under incomplete information, is exactly what case pedagogy uses to build early-career analytical skill. This is the precondition the Layer 2 atrophy paradox turns on: when models already perform the analytical work juniors learn context on, the tasks that once taught discernment stop being assigned to the people who most need them.
-
Lenny Rachitsky, Lenny's Newsletter·evidence
The tech workforce is splitting into an amplified half and a shaken half, and which side a worker lands on predicts sentiment more than title, tenure, or company size.
Asked how AI has shifted how they see themselves as professionals, 49% chose "amplified, I can do more and better", while the rest spread across redefined, destabilised, diminished, or unchanged, and that identity variable predicted optimism, burnout, and layoff fear more strongly than any structural measure in the survey. The productivity-up, quality-worried split runs through the responses: most report measurably higher output yet many fear the gains cost the sharpness of both the work and the worker. That self-reported tension lands directly on the atrophy paradox Layer 2 tracks, without resolving it, since the same introspective self-report reads as augmentation on one side and skill erosion on the other.
-
Anthropic·evidence
Domain expertise, not coding background, predicts who gets useful work out of a coding agent.
Non-engineers with deep domain knowledge matched software engineers' success rate at agentic coding, and expertise rather than job title tracked how much useful work each instruction produced: expert users triggered about 12 agent actions per prompt against 5 for novices.
-
Counts, Zaikin, Cambon & Farach, Microsoft·evidence
Workers ask Copilot for 1.68 work activities and receive 3.26, without perceiving the surplus they delegated.
Microsoft's researchers call the distance between what workers think is happening and what actually is a legibility gap, and argue the unit of accountability has moved from the task to the conversation, since work that never surfaces in the request is harder to audit, attribute or correct.
-
Shen & Tamkin, Anthropic·evidence
Engineers who leaned on an AI assistant to learn a new library scored 17 points lower on comprehension, with the gap widest on debugging.
-
Tobi Lutke, Shopify·illustration
Learning on the shop floor: River works in the open, so debugging and delegation are visible at the organisational scale rather than at the individual one.
-
Microsoft 2026 Work Trend Index·evidence
86% of AI users say they treat AI output as a starting point, not a final answer.
-
Bret Taylor, Sierra·illustration
"I was proud of the elegance of the code I wrote. I haven't quite visualised what replaces that."
- No challenges yet. The atrophy paradox needs counter-evidence to stress-test it: a setting where heavy delegation has not produced skill decay, or where the predicted mid-career hollowing has reversed under specific practices. Actively looking.