Field report · 3 min read
Grex added nine changes to four websites
Guides, comparison pages and a pricing fix went live after a separate AI reviewed every change. Eight other attempts failed.
Grex runs bounded tasks, called missions, that AI agents carry out on computers the operator owns. On October 28, 2026, Grex ran 18 missions on the repository for four live websites: DrivewayGold, HeirLogics, HalfYourStuff and AI-Temize. Nine of the 18 missions were accepted, eight failed and one was cancelled.
How a mission becomes a live change
A builder agent makes the change, and mechanical checks run on the result. A separate reviewer agent then judges the exact change against the acceptance criteria. Accepted changes become pull requests, and the operator merges and deploys them.
The nine accepted changes
Nine changes were accepted, and all nine are now live. Four of them are new guide pages:
- What sells best at garage sales, published on DrivewayGold.
- What an estate sale is, published on HeirLogics.
- Splitting household assets in a divorce, published on HalfYourStuff.
- A printable home inventory template, published on AI-Temize.
A routing fix now makes the DrivewayGold guide load at its address. Three comparison pages were also added, about apps for executors, divorce inventory apps and home inventory apps. The executors comparison went to HeirLogics, the divorce inventory comparison to HalfYourStuff, the home inventory comparison to AI-Temize.
One further change updated the home pages of AI-Temize and HalfYourStuff. Those pages now state that extra credits cost the same $10 price.
Where the comparison pages got their facts
Each comparison page was built only from facts in an accepted research brief. Those facts were quoted from each company’s own website or app store listing.
The one rejection and what it caught
One comparison page was rejected the first time it was reviewed. The reviewer found that two facts in the draft had been restated unfaithfully. One product’s free tools were credited with help the product’s site does not attribute to them.
Export formats were presented as a general feature of a product when the source described only its sample. A correction fixed both problems, and the corrected page was then accepted.
The two models that did the work
The builder was the MiniMax model, and the reviewer was Claude. The reviewer ran on a different machine from the builder.
What failed, and what this report leaves out
Eight missions failed and one was cancelled, so only the nine accepted changes reached the live sites. This report does not cover the reasons those attempts failed or why one was cancelled. It also makes no claim about how the live pages perform after the change. The nine accepted changes are live, and that is the whole of the result.