Sber, Yandex, and red_mad_robot showed how AI is changing the developer's role

AI is increasingly taking over the routine aspects of development, and the engineer's role is shifting from manually writing code to task formulation and result validation. Examples from Sber, Yandex, T-Technologies, and red_mad_robot demonstrate that this shift already impacts not only release velocity but also the competency requirements for developers themselves.

How development is changing

Sber describes the movement toward AI PDLC through a maturity scale spanning zero to level five. The company currently operates at the supervised automation level: AI is embedded in most development stages, but final decisions remain with humans. According to the company, approximately 14,000 developers use GigaCode, with nearly 80% engaging with the tool daily.

The percentage of AI-generated code accepted by teams has grown from 45% in early 2025 to 69% by year-end. This is no longer a point assistant—it's a new work pattern. The next step involves restructuring the environment itself.

In November 2025, Sber transitioned developers from JetBrains IDE to its proprietary GigaIDE PRO, where seven AI agents operate: from documentation and logging to automated testing and analytics. More than half their suggestions are adopted. Simultaneously, hiring economics are shifting: whereas new engineers required 71 days to reach full productivity in 2024, that figure is now 36 days.

In this model, humans increasingly manage the process rather than execute mechanics.

"90% of implementation is performed by AI, while 90% of concept

management remains with people."

What's happening with people

Technological gains do not eliminate human costs. Sber researchers note that junior developers increasingly occupy an awkward position: they must evaluate model-generated results rather than learn to write code from scratch, despite lacking foundational knowledge. Experienced engineers face a different challenge: much satisfaction derived from solving complex problems disappears, and with agents comes the sensation of being a pipeline controller rather than the creator.

This breeds anxiety about skills, status, and market value. At production scale, the tension is sharper. In large corporations, you cannot simply enable agent mode and expect success—systems must understand internal APIs, security policies, and regulatory constraints.

Pure "vibe coding" rarely works for enterprises. GigaCode's recent case illustrates this: in December, they shipped a full tool release for library upload and validation without writing a single manual line of code. The release addressed five major tasks plus fixes, and the team accepted the results without material objections.

This redefines engineer specialization: less manual assembly, more task framing, code review, and architectural ownership.

How the effect is measured

Yandex and T-Technologies demonstrate that measuring impact by lines of code alone is obsolete. At Yandex, roughly 70% of developers regularly employ AI assistants for coding and produce 10–20% more commits on average. Yet code generation represents only part of the work: significant time goes to information search, design, debugging, and review. The internal AI Chat markedly reduced wiki visits, while DeepAgent, per company estimates, delivers tenfold acceleration on complex codebase research tasks. T-Technologies, conversely, focus on Developer Experience rather than text volume:

• time from first commit to deployment
• share of focus time and context-switch frequency
• speed of first review
• proportion of flaky tests
• duration of new engineer onboarding

By T-Technologies' internal telemetry, the share of regular AI users rose from 17% to 85% over ten months. Four weeks after first exposure to the assistant, 80% continue using it in the IDE and 75% in the web interface. red_mad_robot pushed further: the company built prototypes for web, iOS, and Android in 48 hours, generated approximately 80,000 lines of code and 208 commits, with the team's role effectively reduced to one AI engineer orchestrating agents. At roughly $27 per prototype, this represents not merely acceleration but a new product economics.

What this means

Developers do not vanish, but their work rapidly ascends to a higher level: from code writing to intent formulation, result validation, and agent management. For organizations, the principal risk is no longer that AI generates too much—it is that business will succeed in accelerating releases while failing to restructure training, metrics, and engineering culture to match the new reality.