[1/2] The New SDLC: From Vibe Coding to Agentic Engineering (Category AI4SDLC)
I read May. whitepaper Addy Osmani, Shubham Saboo and Sokratis Kartakis from Google’s AI Agent Training Series. The document is useful not in fashionable terms, but in collecting stable mental models: the tools change every week, and the frame must survive this change. It doesn’t fit in one post, so there will be two:)
The authors formulate the main thesis as a deep shift in engineering – not a new language or framework, but a transition from writing code to expressing intent, when the system translates this intention into working software. For decades, the developer-machine interface was syntax, and now that interface is intent. According to the whitepaper. (beginning 2026), 85% of professional developers regularly use AI agents 51% - daily, and about 41Some of the new code has already been generated by AI. The figures should be read as an industrial estimate, but the direction is hard to dispute.
Next, the authors do an important thing - they dilute two blurred terms. Vibe coding (Carpathian term, February 2025) This is when you “give yourself to the vaibu”, describe what you want in natural language and accept that the AI gave you, and if you make a mistake, just copy it back to the prompt. Agentic engineering is the disciplined end of the same spectrum, where AI writes implementation within carefully designed constraints, tests, and feedback loops, and a person holds responsibility for architecture, correctness, and quality.
The key point is that the difference is not whether you use AI. The difference is how much structure and verification surrounds its output. Telling a tech executive that “we’re vibe-coded the payment system” is cause for alarm. Saying “we practice agentic engineering, where AI implements under human-designed constraints, and coverage with tests guarantees correctness” is another conversation.
The main difference here is verification. Agentic Engineering has two verification mechanisms. Tests check the deterministic part: the function gives this output to this input Evals check the non-deterministic: whether the agent has chosen the right path of steps, whether the tools, or the final bar of quality Without both, it is always vibe coding, no matter how beautifully the prompts are written.
A separate part of the document is devoted to the fact that context engineering is a real engineering skill. The quality of the AI code depends not on the trickiness of the prompt, but on the quality of the context. Authors identify six types of (instructions, knowledge, memory, examples, tools, guardrails) and divide contexts into two kinds.
- Static context (Always loaded, expensive on tokens)
- Dynamic context (on demand, cheaper) The shift from “prompt engineering” to “context engineering” reflects a simple truth: the question is not “how to fool AI into writing good code,” but “what a new team member needs to know to do a good job, and how to encode that knowledge into a form that AI understands.”
The authors noted that AI compresses SDLC unevenly: implementation shrinks from weeks to hours, but requirements, architecture and verification remain human-paced. The specification becomes a new bottleneck. The authors do not hide an inconvenient fact: with the declared growth of productivity in the surveys, the 25–39The METR study showed that experienced developers spent on a number of tasks. 19% more time due to verification and debugging of AI-output. AI does not remove the work of implementation, but turns it from “write” to “check and direct”.
In the next part, we will talk about what surrounds the model: harness. (Agent = Model + Harness)Feature factory model, conductor/orchestrator roles and agentic engineering economics.
P.S. These are some of my thoughts in my report. State of AI4SDLC at AI Dev Conf. There was a chain of intent → context → plan → tasks → implementation → verification exactly about the fact that the acceleration of coding does not accelerate the entire development, but outweighs the load higher in the process.
#AI #AI4SDLC #VibeCoding #Engineering #Architecture #Management