How AI Will Change Software Development: Organized Programming #92 Materials (Category #AI4SDLC)

I’ve collected the materials from my conversation with Kirill Mokevnin. Episode 92 of Organized Programming was released on September 13, 2026, with me as the guest. We discussed how AI is changing programmers, teams, and IT companies. I shared my experience introducing AI at T-Bank—and why accelerating code writing still leaves the rest of software development to deal with.

If a product manager writes briefs faster, an analyst produces requirements faster, and a developer writes code faster, the queues between them may only grow. The more interesting question is how many people and approvals one task passes through before a user sees the result.

We discussed

🔸 Teams with agents One engineer can cover more stages of the work and reduce handoffs between people. The necessary roles still remain. A compact team is a direction of change under discussion, not an established norm for every company.

🔸 Adoption at scale Shared model access, tools, agent skills, and isolated environments provide the technical foundation. The product area itself still has to change the process: exceptions that made a pilot succeed must become reproducible for everyone else.

🔸 Specifications and plan review Kirill described using these practices even in small projects because agents reduce the cost of writing things down. Quality requirements still need automated checks; a document alone guarantees nothing.

🔸 Knowledge and internal platforms An agent struggles with implicit rules and a corporate fork that behaves differently from the original product. Documentation also serves people outside development, so moving all knowledge into Git changes their work too.

🔸 Three levels of measuring value Whether people use the tool, how much time it frees, and what the company gains after all costs are three different questions. Closing more tasks may simply mean that the team reached a less valuable part of the queue. It needs worthwhile new hypotheses and a clear idea of where to direct the time that was freed.

🔸 Learning and the limits of autonomy A finished function says little about what a junior engineer learned: you need to review the task definition, plan, and understanding of the result. Complex legacy systems pose a similar problem—first establish why the system works this way and who depends on its behavior.

Episode materials 📌 Episode page 📖 When an agent writes the code: what remains engineering — the long-form article prepared for the conversation. 🎬 YouTube recording 📝 Edited conversation recap

If you are already introducing agents into a team, what now delays a useful change the longest on its way to the user?

#AI4SDLC #AI #Agents #Engineering #PlatformEngineering #Management