AI can draft the Dockerfiles, but hitting 80 MB across 12 real services needs a human in the loop.

The pattern of multi-stage builds is well-understood and structurally repeatable, but each of the 12 services likely has unique dependency trees, runtimes, and build toolchains that require individual analysis. The task is not a simple template-stamp operation.

The 80 MB ceiling is a crisp numeric target, but whether it is achievable for each service depends on runtime requirements the agent cannot fully know without building and testing. Success criteria are clear in principle but hard to verify without execution.

The agent needs read/write access to all 12 service repos, the ability to run Docker builds to measure image sizes, CI/CD pipeline configuration access, and knowledge of each service's runtime dependencies — a broad permission surface that is rarely pre-granted.

A broken Dockerfile or an overly stripped distroless image can silently remove required runtime libraries, causing production service failures that may not surface until deployment. Reverting 12 services simultaneously is operationally painful.

Deciding which dependencies are truly optional, whether distroless is appropriate for each service's debugging and observability needs, and how to handle services that cannot realistically hit 80 MB requires engineering judgment that goes beyond pattern matching.