A detailed analysis of Sber's massive whitepaper 'AI-Disrupt PDLC' has been published on Habr.

A detailed analysis of Sber's massive whitepaper 'AI-Disrupt PDLC' has been published on Habr. An AI enthusiast sifted through 337,000 characters of the document and extracted the real architecture for transitioning to agent-based development, filtering out marketing for C-level. The main focus of the review is on the paradigm shift: from writing code to forming intent, where code becomes a secondary artifact and specification is primary. The author pays special attention to the concept of the Discovery Gap. He emphasizes that simply adapting the old conveyor to new tools gives a linear ceiling of 11-25% growth. In terms of economics and security, the author notes that multi-agent architectures consume about 15 times more tokens than the classic chat mode, which requires the mandatory implementation of FinOps safeguards (Cost circuit breakers) to prevent looping. The telemetry analysis buries manual confirmations of actions (Human-in-the-loop): in 93% of cases, engineers approve requests automatically without reading them. Instead, a transition to batch approvals, trust windows, and an adaptive autonomy ladder (R0-R5) is proposed. As the author of the review notes, Sber is currently at the 3rd level of maturity out of 5 (Supervised automation). Meanwhile, developers have already transitioned to their own GigaIDE PRO, and the share of accepted AI code through GigaCode has reached 69%. Among the cultural side effects highlighted in the review is the 'junior paradox' (newcomers are forced to review complex code that they cannot yet write from scratch) and the changing role of seniors, who lose dopamine from independently solving complex tasks.