A proposed change reached a real website

A Grex agent proposed purchase prompts for an estate-inventory website. The engineering team reviewed the proposal, corrected one detail, and deployed it. An independent readback confirmed the change was live.

That was a real change originating in Grex’s work. It was not an autonomous deployment, and there was no evidence yet that it improved sales.

Grex coordinates AI teams under missions with targets, budgets, and checks. Today’s missions were mostly rehearsals: limited trials designed to expose defects before larger assignments.

Three early rehearsals failed in different ways

One planner was asked to edit existing pages and proposed new ones instead. Another copied a broken example twice before its mission stopped a third attempt.

A social-reply drafter returned five empty responses. Its model reached the output limit before finishing the answer.

Follow-up trials confirmed repairs to both planning problems. The second repair also exposed a publication rule that demanded an unsupported revenue estimate; that rule was corrected.

The drafting repair passed a live rehearsal in the evening. It produced four independently verified reply briefs against a target of two, with $0 of recorded external spend. These were drafts for review, not replies posted to social media.

One useful website proposal, with assistance disclosed

The purchase prompts were added to HeirLogics’ probate guides. A person reviewed and applied the agent’s recommendation, and the live page was checked afterward.

No sessions had reached the change when it was assessed. Its effect on conversion was therefore unknown.

Other website changes also went live that day, but only one of four was Grex-proposed. A separate result required a person to supply the next instruction after a planner repeatedly proposed an already-published action. That result was labeled assisted.

A cautious market brief was also useful

A local lead-generation assessment recommended holding rather than acting. It omitted a revenue estimate because the available search-volume figure was too uncertain to support one.

Independent verification confirmed the business and the brief’s other claims. Leaving a field blank was appropriate when its evidence was missing.

Later attempts still found defects

The repaired growth drafter produced two proposals without truncation. Both failed review: one repeated purchase prompts already live, and the other invented a $19 product.

Repairs were built, but failed mission simulations kept the release on hold.

An overnight social mission then reported its target met at zero of twenty-five and skipped every feed. It was stopped after repeated evidence of the defect, at $0 recorded spend.

The day produced twelve filed defects and $0.36 in external spend. Model usage and oversight effort were additional costs. The strongest outcome remained narrow: one Grex-proposed change was live and checked, with its business effect still unmeasured.