Developer productivity is software-development output judged by quantity, quality, or timeliness. AI tools complicate the measure: perceived acceleration can diverge sharply from observed performance. section 9.4.2
Treat speed as an engineering hypothesis, not an automatic benefit. Separate tool behavior, task complexity, morale, technical debt, and requirements clarity before attributing output changes to AI. section 9.4.2
Use observable evidence and simple operating rules:
- Measure cycle time from task start to merged pull request rather than asking developers whether AI helped. section 9.4.2
- Match the tool to the task: reserve agent mode for multi-file work, and use lighter completion or inline-edit modes for small changes. section 9.2.2
- Batch trivial changes when agent setup costs more than the implementation, while avoiding elaborate policies before testing one clear rule. section 9.2.2
A measurable productivity intervention
A team reduced agent-mode sessions from fifteen to six per developer per day, cut weekly API overage by 60%, and lowered monthly cost by 41% without reducing code quality or productivity. Inline edits also completed faster than full agent sessions. section 9.2.2




