A Video Cites Its First Real Finding

Grex assigns bounded tasks called missions to teams of AI agents, each with its own budget, time window, and target result that Grex checks before trusting it. Today mixed a real milestone — Grex’s first short-form video built from a genuine finding — with an unresolved thread that occupied engineers most of the day: a growth-campaign pipeline that closed one bug only to hit a different one on the same page.

Overnight outage, then a fast fix and an honest empty result

The machine that runs Grex’s opportunity-discovery missions went unreachable overnight, the same host-reboot pattern seen before, and had been down about eight hours by this morning. It came back on its own between roughly 6:00 and 6:35 a.m., no fresh mission launch needed.

Mid-morning, a discovery mission failed for the third time on the same defect: its own strategist component believed it had published a finding, while the system counting results showed zero. Engineers traced and fixed it in about 13 minutes, adding typed counters that track a publication through every pipeline stage so the claim and the record can’t drift apart. The fix shipped within the hour.

A rehearsal — a bounded dry run proving a pipeline works before committing real budget — then succeeded fully: three of three target findings independently verified. That clean result launched a real business run later in the day, spending real budget against a real target. That run came back empty at its three-hour window, zero verified findings, and reported so honestly rather than manufacturing a result. An honest empty result is not a delivered brief: the mechanism is proven, the deliverable still isn’t there.

A second real video, and a QA catch worth noting

Grex’s video pipeline — an AI team that builds short demo clips from a completed mission’s verified evidence, with a critique step checking whether each shot actually shows what its narration claims — produced its second verified video today. For the first time it cited a real audit finding: unlabeled form controls and a missing primary action button on a road-trip weather-planning site’s map page, found by an earlier website-audit run. The critique step caught one shot mismatch — footage labeled a “mission dashboard” was actually a single mission’s detail view — a genuine catch, not a rubber stamp. The video still verified overall and goes out today as Grex’s first real short-form video post.

A false alarm on a second site

The website-audit mission type, which uses browser automation to check Grex’s own sites for usability problems, ran again on a second small site and reported a high-severity page-down finding. A manual check showed all five pages returning normal responses — a false alarm. The real cause, in the mission’s own logs: too many missions running on the same machine at once contended for a single browser session, the automation tool failed to load pages, and that tooling failure got miscounted as a real outage. It’s a sibling of an earlier bug (a missing browser installation, same wrong classification) but a different trigger the earlier fix didn’t cover. Filed to engineering, unfixed as of this evening — and, correctly, not forwarded to the list of changes to apply to any live site.

One bug closes, a different one opens

Grex’s growth-campaign pipeline proposes small, evidence-backed page changes on the small-business sites it manages, then reads back the effect after a change ships. For days it had been trying to propose a genuinely new explainer page on the same weather-planning site, and kept getting blocked by the safeguard meant to stop it from re-proposing already-completed work — because the new page was bundled with one already-done change, and the whole bundle was thrown out instead of just the finished part.

Engineers traced this across several attempts today, from confirming the drafting step’s output to diagnostic tracing to the actual cause: the suppression check compared whole bundles instead of individual proposed changes. A late-day fix now filters out only the completed pieces of a bundle. A fresh test confirmed it: the new page proposal passed suppression cleanly for the first time. That same test then hit a new problem — the drafting model’s output didn’t parse into a valid proposal, despite a complete response. That failure was filed to engineering, unresolved at close; the operator stopped the run rather than leave it unsupervised overnight. One bug class is closed and confirmed, but growth-campaign work on this site is blocked again, by something new, into tomorrow.

The rest of the day, and the money

A social-listening mission, which drafts replies to relevant public conversations, used its one daily retry and still missed its target — an ordinary quality miss consistent with recent days, not a new defect. No new social business run launched, since that only happens after a clean pass.

Revenue is still $0 across the portfolio, with no new purchase or signup events. Measured external spend today was $0.00 — that’s what the missions metered, not the actual cost of compute, subscriptions, and engineering time the day took. Two defects carry into tomorrow: the growth-campaign parsing failure, and the website-audit false positive under concurrency, whose finding was correctly withheld from ever reaching a live site.