The compiler catches the first class of failure.
Types, syntax, and imports are necessary checks. They are not evidence that a user can see, understand, or use the result.
Field Note 05Audience bridgeApril 2, 2026
Public thesis
Vibe Coding Broke
Trusting the output is not the same as verifying the work.
Agentic engineering is not a better prompt style. It is a trust architecture with verification at every level.
Read the original on XThe evidence card
A concrete break is more useful than a general warning.
This is the most accessible doorway into the larger body of work. It takes three failures from the archive and explains why the transition away from vibe coding is an infrastructure shift: the system must prove what the agent cannot see.
3verification layers named
27postmortems behind the thesis
198agents in the operating data
0guarantee supplied by a green exit code
The model
The verification stack.
The visual is an editorial compression of the source essay: enough structure to orient a reader, without pretending that the model replaces the underlying evidence.
01StructuralCompiler, typecheck, imports
02VisualBrowser, routes, screenshots
03SystemicHooks, budgets, cross-file signals
The argument
What changed when the break became visible.
The site version keeps the source's operational insight while making its sequence easier to inspect, share, and connect to the rest of the archive.
The browser catches what the compiler cannot.
Real routes, DOM counts, screenshots, and visual gates force the agent to prove that a feature exists outside the source tree.
Hooks catch what no single file review finds.
Cross-file anti-patterns, accumulated performance debt, and architectural drift require signals that persist across edits and agents.
The boundary
Useful systems name what they cannot prove.
Verification is infrastructure, not a prompt.
Limit
The article bridges a broad audience into a specific operating claim; the claim still needs project context and evidence to be applied responsibly.