The Codex app illustrates the shift left of IDEs and coding GUIs

Reading Code vs “Vibe Coding”

• A core thread debates the claim “I don’t read code anymore.” Many argue understanding software still fundamentally requires reading and reasoning about code, especially for debugging.
• Others say they now mostly interact with agents (Claude, Codex, Gemini, etc.), using them to read and modify code, and only occasionally inspect snippets.
• Several commenters report that just reading code rarely suffices; they need to run, probe, and iteratively modify it to build understanding—whether written by humans or AI.

Specs, Shift Left, and “Waterfall 2.0”

• The article’s “shift left” framing is interpreted by many as a return to spec‑driven or even waterfall-style development: heavy up‑front requirements and architecture, then agents generate code.
• Some see this as the right direction: future engineers will write specs, harnesses, and tests, while agents handle implementation details.
• Others argue real projects evolve after deployment; the code often is the spec, making rigid prewritten specs unrealistic.

Quality, Technical Debt, and Safety

• There’s broad concern that “vibe coded” systems are already fragile: superficial architecture, poor security, bad performance, and weak UX.
• Many predict massive technical debt and frequent rewrites in 2–3 years, though some counter that we already throw away lots of human-written systems.
• Safety‑critical and high‑reliability domains are repeatedly cited as places where “don’t read the code” is unacceptable.

Black Boxes, Testing, and Verification Bottlenecks

• Critics see managing agents without reading code as embracing a black box whose inner workings are neither reproducible nor accountable.
• Proponents respond that the real output is behavior, not source, and that effort should move to specs, tests, static analysis, and “testing ladders” from unit to e2e.
• Counterpoint: if AI writes both code and tests, misinterpreted intent may be faithfully enshrined rather than detected. Verification, not generation, becomes the bottleneck.

Who Benefits and Changing Roles

• Several note AI coding particularly empowers non‑traditional or non‑expert programmers to build internal tools and small apps they never could before.
• Others worry this accelerates production of low-quality “AI slop” and that experienced engineers will increasingly act as “AI janitors.”
• A recurring view: future roles bifurcate into spec‑/agent‑managers vs deep system engineers who still read and craft code carefully.

Skepticism About the Author and Codex App

• Multiple comments question the depth of the author’s engineering background and frame the post as influencer/consulting marketing.
• Some find the Codex app’s UX and feature depth underwhelming compared to mature IDEs with AI plugins, and see its “shift left” framing as overclaiming where the industry actually is.