- Author
- Grace Kind
- Published
- 3 September 2025
- Kind
- Short personal blog post, with one demonstration page
- Original
- gracekind.net/blog/browsersensorium/
- Subject
- Harnesses and agent loops
What it is
Grace Kind writes a personal blog. The post asks whether the gap between visual and text-based agents can be bridged.
What it says
- The post shows one page that ChatGPT agent and Claude Code see differently.
- The test page hides its HTML with CSS and draws a not-found message with JavaScript.
- The post says a text-based agent may miss how a page looks to a person.
- The post says a visual agent may miss the internal workings of a site.
- The post imagines an agent that flags hidden HTML text as a possible injection.
- It adds console logs, network requests and DOM changes, calling the set a sensorium.
- The post concedes such an agent may be unnecessary or hard to build.
What evidence it gives
The only evidence is one demonstration page and screenshots of two agents reading it. The rest is argument and speculation. No agent was built or tested.
Interests and limits
The post is a short personal essay, and the author has no stated stake. Only two agents were tried on one page. The screenshots were not viewed, so the text describes them from the post's wording. The post is dated 3 September 2025 and shows a later last-updated date.