AI in SDLC: from assistants to agents

The talk continues a summer one about integrating AI into the development lifecycle of a large company: some material is reused, but most is new, and Alexander suggests watching them in order. He starts with assistants: the first mass-scale copilot was GitHub's tool — GitHub having been acquired by Microsoft and now integrated even more tightly into its AI ecosystem — while the most hyped developer tool became Cursor, an environment packed with assistants and even with the beginnings of an agent mode inside it. Agents are the next stage: systems that plan tasks themselves and carry them out without step-by-step human instructions, with the human saying what to do next and either checking the result or relying on the agent. In 2025 that means Claude Code from Anthropic and Codex from OpenAI, which installs as an IDE plugin, works in the browser and also comes as a console tool.

None of this would have moved without the infrastructure around it. Anthropic's MCP protocol, released — in the speaker's own careful phrasing, "I think in November" — late last year, replaced point-to-point integrations with a standard way of letting a large language model call external tools. Half a year later it had been adopted by more or less everyone, and Google shipped its agent-to-agent protocol: once an agent can call tools, agents start talking to each other and need rules of the game. That protocol is worked out a bit better, some of MCP's pains are closed, and the two coexist happily; Google prepared it in collaboration with a large number of companies. Alexander then explains the volume of agent talk through economics. Since 2023 billions of dollars have flowed into AI, both from model providers and from consumer companies improving their development with these tools; and when you put in that kind of money, you expect a return on the investment.

The second pressure is on efficiency: abroad, many companies started cutting engineers after covid without giving up their growth plans, and a mantra appeared that tools would keep the pace from dropping, or even speed it up. The third factor is provider promises: more than six months ago Dario at Anthropic promised that by now 90% of code would be written by AI tools, and that has not happened. Top managers at consumer companies are aiming at exactly these effects, and investors are waiting for what was promised. Alexander asks aloud whether there is a real expectation gap; with a smart approach, he argues, the results can be spectacular. His example is the AlphaEvolve white paper Google published in May 2025: an agentic system for algorithmic and scientific problems. The pipeline is assembled from prompt and configuration templates, an ensemble of models to choose from at each step, a way of evaluating how well the code works, and a database of program versions with their results, which are mutated in the hope of improvement.

The problems chosen were ones where evaluation is easy: scheduling resources in Borg, Google's orchestrator of internal workloads, and matrix multiplication. Scheduling became 0.7% more efficient, model training sped up by 1%, and the TPU design was optimised along the way; Google has been at this for over ten years — first AlphaGo, which knocked out the best Go players in 2016–2017, then AlphaTensor, now AlphaEvolve. The second study, from Stanford, sorts tasks by level of agency: level one means an agent closes the activity end to end, level five that a human still steers the process. Fifty experts assigned those levels across a database of US occupations, while more than a thousand real workers — by the speaker's recollection — were asked whether they would like such work delegated to AI. Out came a green zone, where automation is both wanted and possible; a red one, where it is possible but unwanted, and economics will push it through regardless; and a zone where people would delegate but tools cannot cope, which the authors name as the target for R&D and startups.

Many such tasks sit in well-formalisable areas, software development and administration among them — the disruption software inflicted on other industries has reached the making of software itself. Another paper Alexander reviewed, Software Engineering 3.0 from several universities and Huawei, describes three stages: the code-first era with standard IDEs and CI/CD, the flowering of AI-assisted development with copilots, and an intent-based stage where the engineer discusses the task with a model and hands implementation to the AI. One cannot stop at the "agent did it end to end" scenario, though: the whole development ecosystem matters. A Google white paper on developer goals helps here — goals cover the entire SDLC, are not tied to one tool (write high-quality code, design a solution, find information) and link to the internal developer platform's tools, which yields a journey map, effectively Jobs to Be Done for a development platform. At T-Bank that platform is Spirit, now being made AI-native: lifecycle stages were split by profession — testers, developers, analysts, SRE engineers — with working groups around them hunting for the jobs that hurt most; a job is first augmented with an assistant, and once that assistant works too well the team moves to agent mode.

The choices follow from measurement: mobile and web teams spend 40% of their time writing code, 15% each on review with merge requests and on tests, and 12% each in Jira and the wiki. Agent mode got a large boost once a 480B model was rolled out. In the demos an agent writes a five-letter word game in Python end to end; in Helicopter, a Zeppelin fork serving thousands of analysts, it generates and runs code with a chart; in the test management system it drafts a test case from a description of business logic; and a cover agent writes unit tests and sends a merge request, its gains obeying diminishing returns — some forty points from about 10% coverage, a much less significant uplift for a mature team. Telling belief in a tool apart from its effect takes measurement: Alexander walks through DORA, SPACE, DevEx and the DX platform, whose recent framework for assistants and agents rests on three pillars — utilisation, impact and cost-benefit analysis. Weekly retention for two tools stands at 79% and 70%; a survey on AI in Russian development runs to the end of November, with a report promised for February 2026.