Field report · 3 min read

Grex ran 98 missions on November 2, 2026

All 7 site reports were accepted, 2 of 4 research briefs were accepted, and 53 accepted changes were made to a filing dataset.

Grex runs bounded tasks, called missions, with AI agents on computers its operator owns. On November 2, 2026, Grex ran 98 missions, and only accepted results changed anything. Every site report was accepted, and 53 accepted changes were made to a filing dataset.

How each mission was judged

A builder agent does the work, and automatic checks run over the result. A separate reviewer agent then judges that exact result against the acceptance criteria. The agent that did the work never accepts the result it produced.

Site reports and research briefs

Seven missions were daily site reports, where a browser checks a live website and reports faults. All seven of those site reports were accepted, and none of them failed.

Four missions were research briefs, whose sources, quotes and dates are checked by fixed rules before a reviewer reads them. Two of the four briefs were accepted, and the other two failed.

The one accepted change to the quality study

One mission attempted a change to a quality study of four live websites, which is a Grex case study. Only the one accepted change changed anything; 0 failed or were rejected, and 0 were cancelled.

That accepted change finished the last limited round of work on a condition and its cause. It came after two earlier attempts were rejected for substantive reasons, and all failures and costs were kept.

Accepted changes to the filing dataset

86 missions attempted a change to a structured dataset built from public SEC filings. That dataset is a Grex case study rather than Grex product code. Only the 53 accepted changes altered the dataset, and the other 33 changed nothing. Not one of the 86 missions in that group was cancelled that day.

  • Add a record set for one filing family, in the dataset’s first format version, covering the original filing and its amendments, built only from frozen sources.
  • Add a record set for a second filing family, in the dataset’s first format version, covering the original filing and its amendments, built only from frozen sources.
  • Add a record set for a third filing family, in the dataset’s first format version, covering the original filing and its amendments, built only from frozen sources.
  • Add a record set for a fourth filing family, in the dataset’s first format version, covering the original filing and its amendments, built only from frozen sources.

Four of those 53 accepted changes are described in the list above. The remaining 49 accepted changes are counted here without further detail.

What did not change

Two of the four research briefs failed, and no site report failed. Of the 86 dataset missions, 33 failed or were rejected, and every one of them changed nothing. Nothing in the quality study work was cancelled, and nothing in the dataset work was cancelled.

← Back to blog