AI Coding Agents Must Cut Maintenance Costs to Avoid Long-Term Productivity Trap
James Shore argues that AI coding agents, which boost code output speed, must proportionally reduce maintenance costs or they will lead to a long-term productivity decline. Using a model based on crowd-sourced maintenance estimates (10 days per month of code in the first year, 5 days per year thereafter), Shore shows that even a doubling of output with unchanged maintenance costs results in productivity falling below baseline after about five months. To avoid permanent indenture to AI-generated code, the maintenance cost per unit of code must be reduced by the inverse of the speed gain: if output doubles, maintenance costs must halve. Shore warns that stopping AI use does not remove the accumulated maintenance burden, and current evidence suggests AI increases rather than decreases maintenance costs. He encourages teams to focus equally on maintenance cost reduction alongside speed improvements.
Key facts
- AI coding agents must reduce maintenance costs in proportion to speed gains to avoid long-term productivity loss.
- Crowd-sourced estimates: 10 days maintenance per month of code in first year, 5 days per year thereafter.
- Doubling output with unchanged maintenance costs quadruples total maintenance burden.
- Productivity falls below baseline after about five months of using a doubling agent with doubled maintenance costs.
- Stopping AI use does not eliminate accumulated maintenance costs from AI-generated code.
- To maintain productivity, maintenance cost per unit of code must be halved if output doubles.
- Current evidence suggests AI coding agents increase maintenance costs, not decrease them.
- Shore provides a spreadsheet model for teams to test different assumptions.
- The article is not anti-AI but emphasizes balancing speed with maintainability.
- Shore is a consultant specializing in late-stage startups with productivity problems.
Entities
Institutions
- James Shore