AI-Assisted Engineering AMA

In the second session with Aleksey Litvinov, viewer questions converge on one issue: what remains engineering when code generation is automated. The discussion spans fast-changing models and harnesses, organizational adoption, verification, and accountability for the product.

Litvinov separates knowledge by lifetime: model versions and settings are volatile, repository rules change less often, while intent, scope, invariants, acceptance criteria, and feedback form the foundation. A universal in-house harness ages quickly; preserving institutional knowledge matters more.

A mandate that everyone use AI does not change practice. Adoption needs platform capabilities from leadership, local AI champions, and middle managers whose workload the tool reduces. A pilot should serve a business goal and measure the whole flow: acceleration is useless when the queue moves to requirements or approval.

Done becomes observable behavior plus evidence. Teams establish a baseline, intent, exclusions, and criteria, move mechanical checks into hooks and CI, and account for blast radius and rollback. One agent should not both make and certify a change; authorship is separate from accountability, and a named owner remains responsible for the pipeline.