News

More Code, More Bugs: Faros Report Finds Tradeoffs In AI-Driven Software Development

• By John K. Waters
• April 22, 2026

Faros, a software engineering data company, says a review of two years of telemetry from 22,000 developers and more than 4,000 teams found that AI coding tools are increasing software output but are also linked to more bugs, more incidents, and longer review cycles, as engineering organizations struggle to absorb a surge in machine-generated code. The report describes the pattern as an “Acceleration Whiplash.”

The 29-page report, dated March 2026, draws on data aggregated from task management systems, IDEs, static analysis tools, CI/CD pipelines, version control systems, incident management systems, and HR metadata. Faros said it focused on teams as they crossed a 50% weekly active-user threshold for AI tools, including GitHub Copilot, Cursor, Windsurf, Claude Code, and autonomous agents integrated into the software development life cycle. The analysis compared each team’s quarters of lowest AI adoption with its quarters of highest adoption and reported only statistically significant correlations.

The company found that task completion per developer rose 34%, while epics completed per developer rose 66%. Tasks involving code rose 210%, suggesting AI tools are having their strongest effect on code production rather than on broader project work.

At the same time, quality measures deteriorated. Bugs per developer rose 54%, the incidents-to-pull-request ratio more than tripled, and median review time increased fivefold. It also found that 31.3% more pull requests were merged without any review, which the report presents as evidence that human review systems are being overwhelmed by higher code volumes.

The report argues that the shift is no longer mainly about AI as an assistant, because 60% of developers now use at least one AI tool weekly, 80% of teams exceed the 50% weekly active-user threshold, and 60% of AI-generated code is being accepted into codebases, up from 20% in Faros’s prior dataset. About 25% of pull requests are reviewed by AI agents, while less than 1% are opened by agents independently.

Faros also found evidence that faster coding is associated with greater cognitive strain. Daily pull request contexts per developer rose 67.4%, daily task contexts rose 17.7%, work restarts rose 13.8%, and the number of in-progress tasks with no pull request or activity for at least seven days increased 26%. The report said the pattern points to an environment where starting work has become easier, but finishing it has become harder.

Regarding code structure, Faros reported that the average pull request size rose by 51.3%, the average files edited per pull request rose by 59.7%, and the average files touched per developer per month rose by 149.9%. It said larger changes are harder to review, harder to roll back, and more likely to widen the impact of failures.

In delivery pipelines, workflow friction intensified downstream. It reported that time spent in progress increased by 225.2%, while lead time from commit to production deployment rose by 480.4% among organizations that instrumented this metric. In a smaller subset representing about 10% of the dataset, deployments per week fell by 11.7%, indicating that more code was being written and merged without reaching production faster.

One of the report’s more pointed conclusions is that stronger engineering foundations did not shield teams from the effects. Organizations with mature DevOps practices and high pre-AI performance showed the same downstream deterioration as other groups, a finding it contrasted with DORA’s 2025 survey-based work, which concluded that AI tends to amplify existing strengths and weaknesses. Faros argues that telemetry captures operational stress more directly than developer sentiment surveys do.

The report also cautions that full agentic software development remains limited in its dataset. Because fewer than 1% of pull requests are opened autonomously by AI agents, Faros said the data mostly reflects a world in which humans remain in the loop while AI serves as a primary authoring tool. It said removing the human reviewer entirely would likely intensify the pressures it identified.

Faros frames its findings as a warning to engineering leaders weighing staffing, tooling, and process changes around AI adoption. Its recommendations include tighter control of pull request size, earlier quality gates in development environments, and using workflow-stage timing as a continuous feedback signal, rather than relying on more review at the end of the process.

The report does not claim to be a longitudinal study of the same teams over time. It says its 2025 and 2026 reports are independent cross-sections with different datasets and that comparisons to the earlier report are intended to show directional consistency, not precise year-over-year changes.