AI Yesterday, Today, Tomorrow

Rafael is responsible at Sber for transforming the production process and for the engineering tools — GigaCode, GigaIDE and everything around them. The path: for the whole of last year and the year before they built an assisted approach — code completion, chat over code, assistants across the process — and towards the end of last year an agent in the CLI and an agent in the desktop app performed very well, which made agentising development the top priority. Making that work meant building context management, policy as code, getting the internal regulators working properly and reassembling the process itself. Sber now watches analytics on how developer skills show up against a maturity model: whom to train, whom to reward and whom to part ways with. Adjacent roles are next — refreshed agents for testing and analysis ship now, their metrics get measured in the second quarter, training starts in the third, and by the third and fourth quarters they reach devops and design, which they reckon will yield hardest.

Alexey is CTO of VK Tech: clouds, a workspace with mail and calendar, financial products and an AI platform of their own. They are doing a broad sowing across the whole chain from an idea to a release at the customer's site. Analysts and product people got a tool with the long-known requirements about requirements baked in — solve one problem, do not conflict — and the effect was measured simply: far fewer tickets come back from development with the verdict that nobody can even estimate this. Developers have installed Claude Code and go to internally deployed models, while Codex, he says, has few users. The same platform — running models inside the perimeter, filtering personal data, logging for audit — is shipped to customers, and the difficulty there is administration rather than hardware. The weak spot is devops: letting an agent near production access is frightening, even with a local model.

Andrey Kulishov leads development of the Sourcecraft Security part and builds Sourcecraft, Yandex's answer to GitHub. He spent his career on compilers and code analysis, and the old pain of that field — triaging static analyser findings, for which whole teams used to invent heuristics — is now handled by asking a model in plain human language; automatic review and the commit and pull request descriptions he always hated appear instantly. His team writes no agent of its own and takes OpenCode, which Andrey calls the only genuinely open, full-fledged agent. His numbers are sober: per a talk by Andrey Popov, who measures productivity at Yandex, 80% of developers touch AI, but only about 30% use it daily — he quotes 26 at one point and 36 at another — while commit count grows 10–20%. He is sceptical about those figures and dreads to think what happens in companies with neither model access nor money for a subscription.

Alexander recalls how the metrics evolved: first Lines of Code, then AI Lines of Code, and now a sense that someone is chasing maximum token consumption on a burn-it-all principle. Alexey counts context handoffs and the number of people involved in an epic — some stages now merge — counts how often requirements bounce back, and plugs models into hard incidents to measure time to resolve there. The most convincing comparison sits inside a single team: an epic run the classic way with roles split versus an epic run on the fast track, and the difference is enormous, even though the share of people involved with AI, counting more than developers, is by his reckoning under 30% and perhaps 15%. Rafael admits they watch token consumption whether they like it or not: it is convenient for pointing a finger, though it is unclear who the good one is, the person spending a lot or a little. Skills matter more to him: his talk laid out six maturity levels, and the top consumers of inference are often people who assembled automations and burn tokens for nothing.

To the business Sber shows other things: stories reaching production per head, and the effect on time to market. A 60% growth in pull requests and commits last year drew a question from the board: perhaps you invented work for yourselves, and there are no more features in my products? The answer was GigaView, which uses a model to estimate the effort behind written code and compares it with a benchmark — how a notional mid-level developer would have written that code in ideal conditions, with sound requirements and five hours of productive time. The result is notional code hours in arbitrary units, guarding against the temptation to split five stories into fifty. The algorithm was checked against ten years of repository data: in the first months of pandemic remote work and in the first months after the SVO began the measured effort always rose and later returned to normal, while steady growth began with models that gained alignment to code, jumping sharply from the middle of last year. Sber now runs into the throughput of the analysis stage, which was expected: if development accelerates radically, teams will push analysis themselves. Alexander likes that pull-system logic, more honest than forcing agents into discovery, where much is subjective.

The new bottleneck is review. Andrey is surprised by the modest 10–20%: his project has developers who genuinely became tenfold, and by the evening a stack of pull requests is waiting. Systems cannot replace review yet — even flagship models will delete a chunk of a test if you lean on them too hard, and he and his colleagues collect such cases in a separate folder, despite having agents.md and all the rules written down. Alexey offers an option: review should be done by a different agent with a different prompt, the classic developer-versus-tester conflict; the unsolved task, he says, is checking what was intended against what was built without sending the product manager to verify by hand. Sber's answer on organisational rules is policy as code, in which cybersecurity, architecture, reliability and risk translate baseline requirements into a machine-readable format, Markdown in essence: agents take them into account at design time, regulators get checking agents of their own (cybersecurity, he says, has agentised some 60% of its checks), manual checks remain for the most critical systems, and generated code passes the same gates as human code. Alexander calls this shift left finally come true.