Agentic coding and persistent returns to expertise

Anthropic has published research indicating that coding agents perform most effectively when they have access to long-context windows and iterative feedback loops. The study suggests that while automated tools can handle routine tasks, human developers remain essential for complex system architecture and high-level problem solving. By analyzing performance patterns, Anthropic concludes that agentic systems are currently best suited as collaborative tools that amplify the skills of experienced programmers rather than replacing them entirely.