Building a human review layer for agent loops: https://xitter.cf/t.co/xpviTwvcis
Granby, Québec
Joined January 2008
- Tweets34.4K
- Following102
- Followers1K
- Likes10.1K
As a fun side-project, I had Claude build charades.lol/ last week-end.
Hope you have some fun with it :)
Loupe is out. It's a review tool for the two things your coding agent hands you and that you rubber-stamp anyway: the plans it writes, and the pages it builds. loupe.ac/
AGPL, and there's full documentation on running it yourself with Docker if that's your thing: github.com/ubermuda/loupe
The landing page is itself a live demo. The review UI and the widget on it are the real ones, go poke at them and tell me what breaks. loupe.ac/
Got tired of skimming 2000-line plans and typing "looks good". So I built Loupe and open sourced it. Annotate your agent's plans, or click an element on the page it built and comment on it. Your agent pulls it all back through MCP. loupe.ac/
I stopped reading the implementation, so something else has to catch the regressions. Turns out it takes five different checks, and every one of them has something it can't see: ubermuda.xyz/posts/swiss-che…
Last slice is me. I read plans closely and diffs selectively. My blind spot is that I'm not reading the implementation, which was the entire point of the exercise.
So it isn't trust, it's arithmetic. Five blind spots that don't overlap, and the habit of asking which layer should have caught the last thing that slipped through: ubermuda.xyz/posts/swiss-che…