State of AI4SDLC: Research, Processes, and the Team Lead's Role

The research grew from an observation: the world has plenty of reports on how AI affects development, Russia has barely any, and the methodology often depends on the vendor who commissioned it and the answers it wants. The questions followed: where AI is used most, where it stays rare, what happens to productivity, quality and trust, and whether any of it turns into results. Two tracks supplied the answers — a meta-study of more than fifty sources and an in-house online survey. Four ways of measuring turned up: surveys of engineers and technical leaders; telemetry, where researchers get access to companies' git and track commit dynamics, as GitHub does; field experiments such as METR's, which gave sixteen developers new to such tools work in large repositories they had maintained for years — the context does not fit, and AI did not help; and internal corporate studies, hard to transfer to the industry. Among the respected series he names DORA, over ten years old and bought by Google Cloud, the Stack Overflow survey and JetBrains research.

Their own survey, by Alexander's reckoning, went out to an audience of roughly eighty thousand engineers: about a thousand began it, fewer than four hundred finished. The questionnaire was modelled on DORA; the causal analysis is unpublished, so the results are descriptive. He states the limits up front: the sample may be unrepresentative, the effect cannot always be attributed to AI, productivity metrics are heterogeneous — DORA, SPACE and DX coexist inside one company — and there is a novelty effect. AI has become the default tool: around 90% of engineers use it in some form. Growth is uneven: code generation, debugging and documentation are widespread, code review and legacy work lagged. Coding accelerates faster than delivery — before 2024 throughput and stability actually sagged, even as every engineer felt more productive. In an MIT report 60% of ideas reach a pilot, 20% of pilots roll out and 5% deliver business results. In their own survey penetration is lower than abroad, quality improved for 30%, worsened for 15%, and about half do not trust the output.

Last year many built agentic scenarios, but they worked patchily and hit the limits of the models; in September new frontier models arrived, quality caught up and the scenarios converged. He calls this the year in which autonomous agents are treated as members of the team. The scheme where every SDLC stage belongs to a job title — product manager, analyst, developer, QA, SRE, support — gives way to one where agents take scenarios end to end: roles remain, but one engineer can cover the engineering ones. Startups show it, corporations find it harder, so agent-based development is piloted on greenfield. At T-Bank they spent last year automating scenarios inside roles — code review, test generation, spec writing, a coding agent — though not fully autonomously. The bottleneck meanwhile moves downstream: generated code has to be reviewed, pushed through tests and the pipeline, and shipped. We are talking DevOps at DevOps Conf, he notes, and the work has grown: coding, in the words of Claude Code's creator Boris Cherny, is solved; how to ship it onward is not.

He looks for confirmation in Western companies. Google broke its development cycle into stages and jobs, measured them and added AI where it hurt; Meta published a piece on ephemeral just-in-time tests that a second agent generates for the pull request, because tests written after the fact cannot keep up with the pace of change; AWS compresses roles into one cycle with guardrails — and publicly ties several incidents to juniors and mid-level engineers accepting the internal tool Kiro's suggestions, such as redeploying the production environment; one such case cost six hours of downtime, and the final sign-off then went to seniors. DORA puts it differently: AI amplifies what an organisation already has, a strong platform or chaos. There is no free acceleration: Claude, Codex and Cursor work on Alexander's projects at once, and then it all has to be merged. Management layers are compressing too: Meta runs one manager to fifty reports, Amazon is raising the share of individual contributors, Microsoft and Google are flattening. Yet managers drown in exceptions, seniors become routers of problems, and everyone burns out.

The fourth part is about tasks. Work starts before the ticket: context is plentiful enough that a task is easier to assemble from it than to type; manual triage does not scale, since tasks created at a click make the backlog grow without end, and statuses cannot be synced by hand while people and agents work in parallel. Linear published an essay in March declaring task management dead and pivoted into a platform for people and agents. Atlassian, which owns both the tasks and the knowledge base, is adding agents and a knowledge layer, but the market does not believe it — its capitalisation fell threefold — and at the end of last year it paid a billion for DX, a platform grown from the DORA and SPACE world. Hence the practical list: tasks created from incident channels and chatops, traceability from intent to deploy and incident, and context quality, where their internal assistant got stuck. Alexander stacks this year's goals as a ladder: adoption and agentic scenarios, the spread across workflows that matter, throughput and cycle time, quality and stability counter-metrics and economics.

There were many questions from the floor. On the product engineer pattern, the one-person team that predates AI: Alexander had not studied it, though he works that way himself; such people used to be rare and barely moved the industry, whereas the same expectation now goes to engineers at large. That reminded him of the internal tools: an assistant called Nessy around code and Jupyter notebooks, and an in-house Lovable equivalent so product people and designers can build a front end themselves. On economics: spending the gains on savings or on new products is a business decision, while the floor under costs depends on whether open-weight models for your own hardware keep being published. On the human limit: cognitive load has jumped — instead of flow you invent tasks, load agents and review the output, and five hours later you get up wrung out; the hope is that models keep growing more autonomous, and Kurzweil no longer reads as fantasy. On rolling out to thousands of engineers: platform teams embed the basic scenarios into the Internal Developer Platform, while on top of an agent platform with sandboxing teams build their own — an agent assembling advertising audiences from a plain-text brief.