AI coding tools are making clear requirements, constraints and tests more important than the act of typing code, software architect Joy Ebertz argued in a new InfoQ podcast. When an agent can produce an implementation quickly, ambiguity becomes the limiting factor.
Ebertz distinguishes casual “vibe coding” for low-risk internal tools from essential systems where developers specify behavior and validate the result carefully. In the second case, the AI resembles a supervised pair programmer rather than an autonomous replacement. Architects must translate business goals into technical constraints while considering security, scalability and long-term maintenance.
Existing systems are harder. Their real requirements may be scattered across code, bug history and institutional memory, making it unclear whether an agent should patch one component or regenerate a larger area. Code review alone may not provide enough assurance when the generated change is large or reviewers no longer understand every implementation detail.
The conversation is practitioner analysis rather than a measured productivity study. Its useful warning is organizational: faster code generation can weaken understanding if teams do not invest in specifications and verification. The industry also lacks a settled way to train junior developers when agents perform many of the tasks that historically built their judgment.