What Mattered
Both sources are from one author, Simon Willison, posted on the same day. Treat them as one practitioner's field notes, not a trend.
Framework-free 3D generation. Claude Opus 5.5 at Medium effort built a 3D scene in Three.js (77KB). In the same thread it ported the scene to vanilla JS and WebGL (98KB). The port runs but looks visibly different from the original, and Willison can't fully explain why. His claim is that the model is "perfectly capable" of 3D work without a framework. That rests on one anecdote, with no benchmark and no performance or maintainability data.
Agent PR descriptions are being ignored. Most changes to his projects (Datasette, LLM) now come in as PRs opened by Claude Code or Codex. The agent-written descriptions are accurate and thorough, but they read as AI output, so reviewers skip them. His fix is a one- or two-sentence opener that a human vouches for. The agent's full write-up goes below it in a collapsed <details><summary> block labeled with the model and effort level ("Claude Opus 5.5 Medium").
Why It Matters
Both posts point the same way: writing code is cheap, and human review attention is now the limit.
The 3D test changes how to judge dependencies. A library mostly saves humans from writing and reading code. If the model writes the code, the main remaining case for the library is that its code is easier to review. The vanilla version is about 27% larger and harder to read. Dropping a dependency cuts supply-chain exposure and build tooling, but the risk moves into code nobody reads. The unexplained differences in the port also mean a model-driven rewrite is a new implementation, not a translation. Any migration plan that assumes "the agent will port it" has to budget for regression testing.
The PR pattern is a free governance control. A named human owns each change's summary, the model that produced it is stated, and the agent's full detail stays available for audit. It needs no platform change. For the same reason, it won't stick without a template: GitHub's editor has no button for it.
The review point matters more for private-AI clients. Teams scaling coding agents will run into reviewer fatigue before they run into model limits. Long, uniform agent output teaches reviewers to rubber-stamp.
Next Move
- Write the PR template. It should have a human-vouched opener, the agent detail collapsed below it, and a required model/effort label. Use it on your own repos first, then offer it to clients as a lightweight way to govern agent work.
- Measure review attention in agent-adoption assessments. Track PRs per reviewer, time to review, and the share of agent-authored PRs merged without a substantive comment.
- Require regression tests for model-generated ports. Any framework-removal or migration proposal should include visual and behavioral checks. Don't promise that a port will match the original.
- Treat the framework-free claim as a hypothesis. Rerun Willison's test on a component that resembles real client work before putting it in a recommendation.