The Pulse
Faros Finds AI Coding Speed Is Moving Bottlenecks Downstream
Faros Research finds that AI coding adoption is helping more software reach production while increasing pressure on review, QA, and incident response. Its latest analysis covers telemetry from 22,000 developers across 4,000 teams.
AI.info Team ·
AI coding tools are helping engineering teams produce and ship more software, but Faros Research says the gains are exposing a different constraint: the work required to verify, repair, and operate what those tools create.
Faros’s latest analysis, based on the most recent 12 months of telemetry from 22,000 developers across 4,000 teams, describes the pattern as a “Speed Trap.” Code creation has accelerated, while review queues, quality assurance, restarts, and incident handling absorb more of the workload downstream. The findings appear in Faros’s latest AI engineering research.
The company says the picture is not uniformly negative. Several effects seen during the first wave of AI adoption are easing, and more work is reaching production. Yet Faros argues that organizations are keeping the pipeline moving partly by accepting larger changes, less human review, and higher operating strain.
AI adoption is moving from assistance to delegation
AI use is already widespread in the Faros dataset. Seventy-nine percent of developers use at least one AI tool each week, while 86% of teams have more than half of their developers using AI weekly. Acceptance of AI-generated code has reached 65%.
The larger change concerns what teams are willing to delegate. Faros says many companies now use AI agents to review between 50% and 80% of pull requests. Autonomous authorship is less common, but agents open 13% to 14% of pull requests at the leading edge of adoption. Teams are therefore more comfortable allowing agents to critique code than to create it, although the boundary is moving.
Larger pull requests are reaching thinner review queues
Average pull-request size is up 71.8% in the latest dataset. Developers are also touching more files and broader sections of their codebases, which increases the amount of context reviewers must reconstruct before approving a change.
Faros links the larger changes to a sharp increase in merges without review. Six months earlier, the company had reported a 31.3% rise in pull requests merged without review. Its latest figure is a 76.3% increase.
The finding does not establish that AI-generated code is responsible for every unreviewed merge. Faros instead describes a capacity problem: output is increasing faster than review resources, and teams under delivery pressure may skip review to prevent queues from growing. The company recommends review requirements tied to the risk and scope of each change.
QA inherits the work that coding no longer absorbs
The clearest downstream shift appears in quality assurance. Time spent in QA is up 300.6%, which Faros identifies as the largest deterioration among the major efficiency and flow measures in the dataset. Review time also remains elevated.
Faros’s explanation is straightforward. Larger and more complex AI-assisted changes arrive faster, while a smaller share receives human review before merging. QA then inherits more of the burden of determining whether the software works as intended. The authoring stage moves faster, but the proof of correctness takes longer.
Developer restarts offer another signal that work is moving further through the system before problems become visible. Restarts are up 66.7%, nearly five times the increase recorded in the prior dataset. Faros describes the new failure pattern as reaching the wrong path and starting over, rather than simply managing too many parallel tasks.
More software ships, but the incident count rises
Faros reports a meaningful improvement in delivery flow. Deployment frequency has reversed direction, task throughput remains strong, and code churn has fallen from the extreme levels recorded during the initial period of AI adoption.
Per-change risk has also stabilized relative to the company’s earlier findings. Incidents per pull request are up 14.5% in the latest dataset, compared with a 242.7% increase in the previous report.
The aggregate operating picture is less reassuring. Monthly incidents are up 125.4%, remediation is slower, and backlogs are growing again. A typical change may be safer than during the first adoption shock, but many more changes are moving through the system. The total cost of operating the software estate can therefore rise even when the risk attached to each individual change improves.
Agentic review shows a possible countermeasure
Faros identifies agentic review as the first intervention it has measured at meaningful scale that is associated with improvements across several downstream metrics. Teams using it heavily see faster first reviews and lower change failure rates.
The company qualifies the result as correlational. Unreviewed merges continue to rise even among teams with high levels of agentic review, so the data does not show that automated review has solved the capacity problem. It does suggest that review agents can absorb part of the workload created by AI-scale output.
Faros’s broader recommendation is to focus earlier in the process: give coding agents better context before they write, rebuild controls around AI-assisted development, and judge AI systems by verified outcomes rather than adoption alone. The immediate evidence for that argument is the queue already visible in the data: QA time up 300.6%, larger pull requests, and a 76.3% rise in merges without review.