State of AI4SDLC on HighLoad++: where development bottlenecks are moving (Category AI4SDLC)

Appeared. record My performance at Saint HighLoad + + 2026, slides and abstract. The report was about what happens when the local acceleration of coding gets into the engineering system of a large company. 10 000+ engineers. The answer is simple: AI speeds up code writing before the organization has time to rebuild the entire supply stream. Therefore, the bottleneck does not disappear, but moves to task setting, review, testing, integration and release. The team can make more changes and deliver value to the user no faster.

Three practical developments follow.

1Moving from role-based SDLC to agent-based SDLC An engineer with agents can close a wider end-to-end scenario rather than outsource work along a long chain of roles. The specification does not return as a heavy document, but as a contract with the agent: purpose, context, limitations, acceptance criteria, and a way to verify the result. Weak problem-setting doesn’t go away – AI just helps scale the error faster.

2Transition from a set of AI tools to an agent-first platform At the scale of a large company, it is not enough to give everyone a good coding assistant and connect dozens of MCP servers to it. You need a model gateway, tool gateway, and feature registry with owners, versions, policies, and quality check kits (evals). Trust must also be designed by step. read -> recommend -> actAutonomy is not given to the agent as a whole, but to a specific action in a particular context.

3Moving from use metrics to result metrics The proportion of AI code says almost nothing about the strength of an engineering system. You need to look at the entire stream: the time before the first merge request, waiting in the pipeline and on the review, alterations, defects, incidents, the cost of tokens and human time. That is, to link implementation with speed, quality, risk and economics.

And here comes an important sociotechnical part. The same tools give teams different effects. The winners are those who change the work itself: turn requirements into verifiable contracts, make context accessible to the machine, embed checks into the process, and leave a clear owner of the result.

#AI #AI4SDLC #Engineering #PlatformEngineering #Management #Conference