Field report · 3 min read

Grex ran 22 missions and accepted all 6 daily site reports

One research brief and several code changes were also accepted, while failed or rejected missions changed nothing.

Grex runs bounded missions with AI agents on computers its operator owns, and the builder never accepts its work. A builder does the work, automated tests run, and a separate reviewer agent judges the exact result against acceptance criteria. On November 10, 2026, Grex ran 22 bounded missions, with mixed results. All 6 daily site reports and the 1 research brief were accepted, and several code changes were also accepted.

Reports covered websites and research

Daily site reports use a browser to inspect a live website and report faults. All 6 daily site reports were accepted, and 0 of them failed. Research briefs have their sources, quotes and dates verified by automated tests before a reviewer reads them. The only research brief was accepted, and its failed count was 0.

Only accepted changes changed anything

Only accepted changes changed anything; failed, rejected and cancelled attempts changed nothing on that day. Every category on November 10, 2026, recorded 0 cancelled missions, and none altered any project.

Accepted changes covered the website and dataset

Grex’s website had 1 attempted change, and only the 1 accepted change changed anything. For this category, 0 failed or were rejected, and 0 were cancelled. Four missions attempted changes to a structured dataset built from public SEC filings. The structured dataset was a Grex case study rather than product code.

Among the 4 attempts, only the 2 accepted ones changed anything at all. The other 2 failed or were rejected and changed nothing, and 0 were cancelled.

Two home inventory experiments were attempted

Five missions attempted changes to a home inventory experiment without a live release. Only the 3 accepted ones changed anything; 2 failed or were rejected and changed nothing, and 0 were cancelled.

Four missions attempted changes to a home inventory experiment about keeping Windows packages separate, without a live release. Only the 1 accepted one changed anything; 3 failed or were rejected and changed nothing, and 0 were cancelled.

The skills storyboard attempt was unsuccessful

One mission attempted a change to the skills storyboard experiment for Linux and Windows. The accepted count was 0, and no accepted change changed the project. One failed or was rejected and changed nothing, and 0 missions were cancelled.

Five accepted changes in detail

Below are 5 accepted changes selected to help a first-time reader understand the day. The remaining 2 accepted changes are counted here but not described individually.

  • For Grex’s own website, the change added only the allowed storyboard file.
  • A dataset change added only the browser test evidence file for contained observations, containing exactly the line “observed”.
  • A home inventory change added toolbar behavior for search, filter, sort, count, clear and no-results, including every required test label.
  • A home inventory change built the specified toolbar with required test labels and the specified behaviors.
  • A Windows change added only the specification file, which contains one automated test confirming one plus one equals two.

Failures changed nothing

Failures and rejections covered 2 dataset attempts, 2 home inventory attempts, 3 Windows package attempts, and 1 skills storyboard attempt. Across every category on November 10, 2026, there were 0 cancelled missions. Only accepted changes altered the website, the dataset, or the projects themselves.

← Back to blog