The largest empirical study to date on AI agents in software development reveals a paradox. Faros AI analyzed two years of telemetry from 22,000 developers across more than 4,000 teams globally. The report’s title says it plainly, “Acceleration Whiplash”: push hard with AI assistance and you will hit a quality wall sooner or later. More output, more failures, less control.
The productivity gains and their cost
Faros data cuts both ways with clarity. High-AI-usage teams complete 66 percent more epics, meaning finished work packages, per developer and handle 34 percent more tasks overall. These are genuine productivity gains from actual enterprise telemetry, not laboratory experiments under controlled conditions.
Simultaneously, production incidents per submitted code batch rise 242.7 percent in the same teams. Average code review time stretches 441.5 percent. Code churn, the share of code deleted shortly after creation, climbs 861 percent. Code gets written faster than it gets thought through.
One more number stands out: pull requests submitted without any human review or automated checks rose 31.3 percent. The report traces this directly: reviewers cannot keep pace with the volume AI agents produce.
When budget outpaces quality
Teams with the highest AI token budget, those leaning hardest on AI-generated code, achieve double the throughput of teams without AI support. At the same time, their token costs run ten times higher. Technical publication ADT Mag captures the core finding this way: more code means more errors if review processes do not scale with production.
AI agents excel at code generation. They struggle to judge context, spot interactions with existing systems, and weigh long-term consequences. This is not a problem future model versions will auto-solve. It is a systems question about who bears responsibility for the outcome.
An OpenAI analysis parsing more than 800,000 professional messages from US companies reveals another dimension: strip out pure routine tasks like writing or scheduling, and 43.5 percent of all specialized work now gets assigned to roles historically disconnected from technical labor. AI agents are pushing engineering tasks into occupational fields far beyond classical developer roles.
Engineers as managers of digital teams
For software developers, this shift redefines the professional role fundamentally. The portal ingenieur.de frames the change this way: engineers increasingly are not those writing code, but those tasking agents, evaluating results, and fixing errors. “What remains for the human?” ingenieur.de asks. The answer: steering, quality checks, accountability for outcomes.
This role shift harbors a structural problem the Faros report makes visible: oversight mechanisms have not kept pace. Who verifies the agent chose the right code for the right problem? Who catches the agent reproducing an existing bug instead of fixing it? Bugs per developer have risen 54 percent by the report’s count. That is the gap between agent output and human verification, expressed as numbers.
The consequence extends beyond development teams. Software gets made by more people in less time and pushed to production with less scrutiny. Every banking application, every medical system, every administrative program built with AI agents carries potentially more undetected failures than before. Whether this quality overhead gets caught depends on how rigorously each team organizes oversight.
Before governance tools close the gap
Gartner values the global market for AI agents in enterprise at 215 billion euros. This growth accelerates not despite the documented quality problems, but alongside them. The market is now developing answers to the problems it creates: monitoring dashboards for agent output, automated quality checks at the PR level, governance frameworks for agent decisions.
When these tools become widely available and production-ready, no one can say. Until then, ingenieur.de states the principle: an AI agent is only ever as good as the person who evaluates and releases its results. The Faros report documents in figures what happens when this principle goes unenforced in live systems: more code, lower quality, higher costs. Whether industry standards or regulatory requirements drive the market’s next move will decide itself over the coming months.
