Simon Willison recently encountered a concrete problem: most of his code changes are submitted by Claude Code or Codex—AI tools capable of autonomously writing and modifying code. The pull request descriptions generated by these agents are so complete that they have reduced his willingness to read them. He now uses HTML collapsible sections to preserve the full text while exposing only a manually selected summary, acknowledging that even accurate information can test a person’s patience.
What this is
The PR (Pull Request) Willison refers to is the basic unit developers use to submit code changes, document the reasons for those changes, and initiate a review. He now places a short, manually curated summary at the beginning of each description, then puts the agent-generated long-form explanation inside a <details><summary> collapsible section. GitHub supports this format natively, allowing readers to expand the full text whenever needed.
This is not a model capability upgrade but a change in interface and editorial method: agents are responsible for producing large volumes of information, while humans decide what people should see first. Willison also asks agents to draft summaries, but the final summary still requires human judgment.
Industry view
We believe this shows that the bottleneck in AI coding tools is beginning to shift from “whether they can write code” to “whether review costs can be reduced.” When an agent submits large volumes of code and documentation at once, developers who cannot quickly identify what changed and where the risks lie may face a heavier verification burden as generation efficiency rises.
But collapsing details does not mean solving the problem. The objection is direct: hiding detailed information may weaken review, especially when the summary is drafted by the same agent and the programmer lacks the ability to verify the underlying changes. Readers may approve the work based on a single generalization. Moreover, Willison’s experience comes from personal open-source projects. Large enterprises also have to account for permissions, auditing, test coverage, and multi-person collaboration; these challenges cannot necessarily be transferred with a single HTML tag.
The responsibility chain deserves greater attention: as code generation accelerates, testing, access control, and human acceptance become more important at the same time. AI expands the scale of output, but it does not automatically expand an organization’s capacity for judgment.
Impact on regular people
For enterprise IT: When introducing coding agents, organizations need to evaluate not only the code they generate, but also the quality of their descriptions, review workflows, and permission boundaries. Writing less long-form documentation does not mean conducting less technical validation.
For individual professionals: The core skill for collaborating with AI will shift from “producing more content” to “selecting priorities and providing a verifiable summary.” People who can explain the reasons for a change, its risks, and how it was validated will be more likely to retain control over review.
For the consumer market: More software will be produced and updated rapidly by agents. Consumers may experience smoother functionality while becoming less able to judge whether automated changes are reliable. Brand credibility will increasingly depend on test records, transparency around changes, and how quickly vendors respond to problems.