- Author
- John Allspaw (Adaptive Capacity Labs)
- Published
- 24 August 2026
- Kind
- Blog essay responding to a research paper
- Original
- www.adaptivecapacitylabs.com/2026/08/24/there-is-more-to-code-review-than-automatable-detection/
- Subject
- Review, taste and expertise
- Themes
- Checking the work, Skill under automation
What it is
John Allspaw responds to a paper titled The End of Code Review. That paper argues coding agents make human review unnecessary. The post says the argument depends on a substitution myth.
What it says
- The paper splits review into four functions and argues agents can do each one.
- A reviewer's confusion is itself a finding, which an LLM cannot supply.
- The post says human reviewers notice what is missing, a weak spot for LLMs.
- Review also questions whether a change is needed, using context from outside the repository.
- Knowledge transfer in review is a joint activity, not a summary an agent delivers.
- Named accountability shapes how a reviewer reads, and agents bear no consequences.
- Review is also coordination, sensemaking and governance, not only detection.
What evidence it gives
The post is argument only. It reasons from how code review works in practice. It gives no data, test or case study. It quotes the paper's abstract and comments on its discussion section.
Interests and limits
The post appears on the Adaptive Capacity Labs blog. It responds to one paper. The paper itself was not read here, only the post's account of it. No agent was tested. Claims about LLM weaknesses are stated without citation.
Connections
These links are this wiki's own reading, not claims the sources make.
- Treat Agent Output Like Compiler Output. 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?