The Cursor Desk
Cursor is one of the harnesses this shop runs on — Desktop sittings on the laptop, cloud agents over the API, a phone for steering. This desk keeps what we know about it: what changed, what it costs, where it breaks.
Most Cursor coverage is written by people reviewing Cursor. This desk is written by a shop that produces with it every day — a fleet of websites, a work-order system, and the measurement rigs behind our citation practice. Everything here rides on our own run record: versions we actually installed, agents we actually launched, work orders that actually came back. Where a claim is Cursor’s own marketing, it is labeled as theirs.
The state of the harness — October 19, 2026
Desktop: version 3.25.9 on Will’s laptop, read from the install itself on October 19, 2026. Desktop is where supervised sittings run — a human present, a work order on the screen.
Code seat: Composer 2.5 — the model every Desktop work order in this shop names as its executor.
Cloud catalog: 46 models on the last full pull (October 16, 2026, via the API’s model list), diffed daily ever since.
Changelog: latest entry dated October 16, 2026 — and a daily changelog watch joins the catalog watch starting today.
Surfaces in daily use: Desktop (sittings) — Cloud API (agent launches from this desk) — Mobile (review and steering; the runs are distilled into our record every 20 minutes).
What this desk measures
Four things, all from the receiving end. Whether work actually landed — the live page, the delivered file, the count re-added by hand. The execution steps behind a result, graded as a trajectory, because summaries are where errors hide. The model catalog, diffed daily, because a default that changes silently changes your work silently. And the changelog, diffed daily, because an upgrade you did not notice is a change you did not choose.
The pieces
Start with the receipts
44 AI models, one identical brief, five comparison rounds — the census that set this shop’s model routing, run entirely through Cursor’s cloud API.
Desktop, Cloud, or API — one harness, three seats
The seat is a routing decision, and it is decided by where the work’s context lives. When each surface wins, when each one failed us, and what each failure changed.
The model catalog, watched daily
A model catalog is a moving part of your stack. Ours is diffed every morning — the October 16 baseline of 46 models is the worked example of why.
Composer 2.5 is our code seat — here is its record
Fastest model on our October 16 board, with a perfect score — and the dents the record also shows. Why it holds the code seat, and why it doesn’t hold the first-pass seat.
What’s new is a claim — here is how we check it
Changelog entries are vendor claims. The desk’s discipline: two daily watches, a read against our own record, and a check before anything changes how we work.
Review the run, not the diff
Cursor’s cloud agents can now hand back a canvas — a rendered board a reviewer opens by link. What the vendor’s own pages establish, the team-only gate on the link, and where a canvas lands in our debrief loop. Our record so far: no canvas yet, stated on the page.
The standing promise
When Cursor changes, this desk changes. The watches fire, the change gets read against our record, and the desk says what it means for a working shop — or says plainly that it doesn’t matter yet. Sister desk: the GLM desk, where the same treatment covers the model family our first pass rides on.