This thesis proposes and evaluates a framework for measuring the productivity impact of agentic coding systems in software engineering teams. The framework consists of five metadata-based metrics covering the two key dimensions of software engineering productivity established in the literature: development velocity and software quality. The metrics rely on metadata extracted from GitHub and Jira. The framework was evaluated through a longitudinal case study of three software engineering teams (team A, team B, and team C) within a Nordic B2B SaaS company over a six-month period. Weekly quantitative data was combined with semi-structured interviews with software engineers from each team and complemented by expert interviews providing context on the broader adoption journey of agentic coding systems.

The three teams exhibited distinctly different adoption patterns. In team A, adoption varied significantly between team members, ranging from non-adopters to engineers who use coding agents for full task execution in most of their work. Team B showed a more uniformly distributed adoption pattern, where all team members use coding agents selectively for tasks they believe agents perform well on. Team C displayed a more foundationally invested pattern, where team members have made an effort to provide agents with the infrastructure needed to complete work in line with the team’s standards. Team C showed the clearest upward trend in both development velocity and software quality indicators, followed by team B, while team A’s individually driven adoption pattern resulted in fluctuating outcomes that could not be reliably attributed to agent usage. Overall, the study shows that adoption is driven by both organisational and individual factors. Software engineers’ perceptions of the advantages that come with using coding agents play a particularly significant role in adoption.

The study concludes that agent usage does not automatically translate into measurable productivity gains. Rather, the quality of adoption, characterised by foundational investment, shared practices, and integration into workflows, appears to support sustained improvement. The proposed metadata-based framework offers a feasible and scalable approach to evaluating the productivity impact of agentic coding systems, but its reliability depends on continuous qualitative input that can contextualise the quantitative trends.