Skip to content
Agentic Engineering Primer

Productivity Research

AI coding productivity results are mixed.

Key Findings

Section titled “Key Findings”
  • METR (2025, updated 2026): Experienced developers were about 19% slower in the tested setup, despite perceiving themselves as faster.
  • DORA 2025: AI tends to amplify existing team strengths and weaknesses rather than fix process issues on its own.
  • Practical takeaway: Gains are strongest when teams have tight feedback loops, clear ownership, and reliable verification.

Why Results Conflict

Section titled “Why Results Conflict”
  • Different studies measure different tasks (greenfield vs maintenance)
  • Tool configuration and model selection vary widely
  • Team maturity changes outcomes as much as model quality

How to Apply This

Section titled “How to Apply This”
  • Run a scoped pilot on your own codebase
  • Measure PR cycle time, review time, and rework rate
  • Track quality outcomes, not just code volume

Sources

Section titled “Sources”