AI can draft the source generator, but a human must own the production swap.

Source generator patterns in C# are well-established and structurally similar across projects, but the specific domain object shapes, edge cases, and endpoint wiring are unique to this codebase. The agent can't rely on a generic template without reading and understanding the actual models.

The goal — eliminate reflection overhead via compile-time serialization — is clear, and '30% CPU reduction' gives a measurable target. However, success criteria around correctness (all edge cases handled), backward compatibility, and which serialization behaviors must be preserved are underspecified.

The agent needs full access to the codebase, the domain object definitions, the three target endpoints, existing serialization configuration, and ideally a benchmarking harness. Without repo access and the ability to run tests, the agent is working blind on the most critical parts.

These endpoints handle customer transaction data — a serialization bug could silently corrupt data, drop fields, or cause runtime failures in production. Mistakes here are not easily reversible and carry real business and compliance risk.

Deciding which serialization behaviors to preserve, how to handle nullable types and polymorphism, whether the performance gain justifies the maintenance complexity, and when the swap is safe to deploy all require engineering judgment that goes beyond code generation.