- Author
- Hugo Venturini, Software Engineer, SkipLabs
- Published
- March 9, 2026
- Kind
- Company blog post, an argument by analogy
- Original
- skiplabs.io/blog/codegen_as_compiler
- Subject
- Compilers and interpreters
- Themes
- Checking the work, Handing over control
What it is
Hugo Venturini, a SkipLabs engineer, wrote this in reply to Philip Su's lights-out codebases post. It is part of a series on verified AI code generation. The author asks why lights-out feels scary and answers with the compiler comparison.
What it says
- Venturini says nobody reads compiler output. Tests, type systems, fuzzing and reproducible builds make that unnecessary.
- The post says agents lack that verification apparatus. The output itself is not what is missing.
- It argues review volume is a symptom, not the problem. One developer landed 417 pull requests in a day.
- The post says quality control moved upstream and downstream, and the manual middle was eliminated.
- It says agents are not deterministic, so test suites must be substantially more comprehensive.
- The post lists four things to build. They are formal specs, test infrastructure, AI-checks-AI pipelines and production instrumentation.
- It says chip makers already verify black-box components by acceptance tests, not human review.
What evidence it gives
The post offers argument by analogy. It cites one developer's 417 pull requests in a day as an example of volume. It names no experiment, benchmark or case study of its own.
Interests and limits
The author works at SkipLabs, which sells tools in this area. The post is part of a company series. Nothing was tested. The bookmarked address redirects to the SkipLabs docs home. The text was read at skiplabs.github.io/skip/blog/codegen_as_compiler.
Connections
These links are this wiki's own reading, not claims the sources make.
- There is more to code review than (automatable) detection. Venturini says nobody reads compiler output, so verified agent code need not be read. Allspaw lists what review does beyond detection: confusion as a finding, and named accountability. Tests can replace detection. Which of the other functions has a compiler equivalent?