How to Use Agents Across the Entire SDLC

AI agents already write code reasonably well, but then they get stuck at "here are the changes, take it from here" — because there's no environment around the agent in which a change can actually be carried through to production. This talk is about how to build such a harness: an engineering loop made of project memory, skills, and MCP that turns the agent from a code generator into a participant in the entire SDLC. Drawing on my experience at Yandex, I'll show how to build the environment an agent needs, for example:

• Project memory — architecture and decisions living next to the code.
• Skills — a shared catalog of the infrastructure skills an agent needs.
• Rules for team collaboration — how to set up skills, hooks, and MCP for a shared project.

We'll also talk about the changing role of the engineer.