Danila Shtan: AI Does Not Replace Fundamentals or Engineering Accountability (#AI4SDLC)
I watched a Beyond Coding episode with Nebius CTO Danila Shtan. The title promises a discussion of the skills that get engineers hired, but I found the conversation broader: hiring, engineering organization, and trust in AI-assisted code all converge on one principle—autonomy does not remove control; it moves control closer to whoever makes the decision and owns the consequences.
In 2021, at HighLoad++, I gave a talk about how the CTO role changes as an organization grows. Shtan’s answer is quite specific: technical depth still matters, but as the organization grows, a CTO primarily builds the team, owns the outcome, resolves friction between engineering and the business, and manages expectations. A commitment is not only something to fulfill; any deviation has to be surfaced before it becomes a surprise in someone else’s plan.
The same pattern appears in engineering skills. According to Shtan, prior AI experience is not required for many AI-cloud problems: the work needs engineers who understand distributed systems, drivers, networking, low-level storage, GPU kernels, and inference optimization. This connects neatly with my recent vLLM and PagedAttention breakdown: a classic virtual-memory idea was used to solve the problem of managing KV-cache memory.
Knowledge of a specific library or framework is moving into the background; fundamental reasoning is not. In the hiring process Shtan describes, coding and algorithmic interviews still happen without AI. More experienced engineers also walk through a recent project end to end: where the requirements came from, who made decisions, why the product was built, what happened after launch, and whether the candidate understood its operational consequences. At the time of recording, the company was only designing a separate agent-coding session. This is a time-bound description from a public conversation, not a promise that the process will stay unchanged.
The same principle becomes stricter at the organizational level. In Shtan’s description, Nebius engineering teams have neither a separate architect nor an architecture committee: each team designs its system, launches it, takes on-call duty, and owns the service. Freedom to decide is paid for with operational accountability. Otherwise, autonomy quickly turns into local optimization at someone else’s expense. Interestingly, I advocated much the same approach throughout my time overseeing the architecture function at T-Bank, although I was often criticized for not having an enterprise architecture department like Sber’s. By the end of my time there, a few large divisions had created their own enterprise architecture teams, but those were local initiatives :))
Shtan draws a similar boundary around AI agents. He compares working with an agent to working with a junior engineer: the agent lacks full context, proposes strange ideas, and may take them a long way. According to Shtan, at the time of recording, employees had access to OpenAI models and Codex, but not broad access to Claude Code. He framed this not as an “AI ban,” but as a gap in guardrails and observability.
His desired protocol for an AI-assisted pull request includes marking the agent’s involvement, preserving the full trajectory of the sessions, making that trajectory auditable, and discussing not only the code but also the original intent: where the human steered and challenged the model. In the episode, this is a target state, not an already-deployed mandatory policy. A change to a production system still has a human owner.
My recent posts on Cursor Cloud Agents, one agent with files replacing complex scaffolding, and Claude Code removing 80% of its system prompt formed a technical story: stronger models need less step-by-step direction, while complexity moves into the environment, tools, state, and verification. Shtan adds the organizational continuation: the more freedom an agent or team receives, the clearer its boundaries and the human ownership of the outcome must become.
My main takeaway: AI makes local code production cheaper, but system ownership more valuable. Task framing, understanding constraints, independent verification, operational context, and accountability for consequences do not disappear. They are exactly what separates autonomy from a lack of control.
#AI4SDLC #AI #Agents #Engineering #Architecture #Management