Confidence: medium. Evidence: production telemetry, behavioral studies, and practitioner counterevidence. Last substantive change: 2026-08.
Delegation changes jobs before it removes them. This area tracks which human roles remain valuable, how teams organize, and how much understanding people keep as agents take on more work.
The conclusion
Expertise remains valuable, especially for planning, domain judgment, exception handling, and governance, and domain expertise rather than coding proficiency is associated with success. The team must preserve enough code and system understanding to challenge the factory, own consequential outcomes, and recover when its abstractions fail. The required depth should rise with blast radius. Higher cognitive offloading is associated with lower task understanding in some studies, but long-term deskilling has not yet been established.
How the thinking got here
The developer role moved from coder to reviewer to specification writer and orchestrator, and then toward domain expert and accountable governor. Alongside that shift came measured cognitive offloading and correlations with lower task understanding, plus concerns about comprehension debt and the burden of coordinating agent work. Large-scale telemetry found that domain expertise still predicts the best outcomes even as people without formal coding backgrounds gain access to technical work. Recent practitioner accounts from Horthy and Willison add a practical counterweight to literal no-read factories: agents pay off when the evidence is strong and human attention is selective, not when people surrender production understanding altogether.
Credible alternatives, and when each is right
| Approach | Right when |
|---|---|
| Fully autonomous cells | low-risk, well-bounded domains |
| Expert-supervised factories | high-stakes or novel work |
| Rotating human immersion | keeping a shared mental model alive |
| Mandatory manual drills | preserving incident readiness |
| Mixed human and agent teams | balancing throughput with understanding |
Where it fails and what we still don't know
Failures may include deskilling, a broken junior pathway, operator fatigue, and teams that lose the shared understanding they need when an abstraction fails. Evidence is strong that domain expertise predicts success. Evidence that delegation causes lasting loss of understanding is not yet strong: current studies find cognitive offloading and associations with lower task understanding, not long-term causal decline. Open questions include skill formation, junior pathways, deskilling and reskilling, and incident readiness.
What would change our mind
Evidence that teams can sustain deep system understanding while delegating almost all implementation would ease the tension between throughput and comprehension.
Evidence and further reading
- Agentic Coding and Persistent Returns to Expertise
- Cognitive Offloading Score
- Reliance and Executive Function Attenuation
- GenAI Productivity and Learning: A Meta-Analysis
- Cognitive Debt: AI as Intellectual Leverage and Systemic Fragility
- Everything We Got Wrong About Research-Plan-Implement
- Engineering Practices That Make Coding Agents Work