Nine failed rehearsals, then one that worked

Grex coordinates small AI worker teams that scan public sources for real business problems, then checks their evidence before trusting the result. One team spent most of today unable to publish anything it found, until a late fix let it launch this week’s first live business run. That run finished on its own hours later, but its result exposed a second gap the evening had to close.

A defect that looked identical nine times in a row

The discovery pack scans places like Hacker News and Y Combinator’s “Request for Startups” page for posts describing a real unmet problem, turning strong candidates into a verified “problem signal” — a source-backed statement of a need someone could build a business around. Before a team can run for real money, it must pass a rehearsal: a supervised dry run proving the pipeline works.

Nine straight rehearsals, across nine pack versions, failed the same way: the pack scored candidates but never published a verified signal. Engineering traced it to a mismatch in how the pack’s writer was told to work — it required source text to state both a problem and an audience explicitly. Real forum posts usually only imply one or the other, so valid candidates kept getting thrown out as too vague.

The tenth attempt, then a second problem

With that contract loosened, the tenth rehearsal succeeded cleanly, no manual help needed. That surfaced a separate gap: the system certifying a rehearsal fit to unlock real spending only accepts one that concluded on its own — a human-stopped rehearsal, even with enough evidence in hand, cannot unlock a run. That’s a known limitation on file with engineering; the workaround was letting the next rehearsal finish itself.

A run that met its target, but not the assignment

That run went live in the day’s final hours and settled on its own about fifteen minutes after close. It met its target of three independently verified and ratified problem signals, spent none of its one-dollar cap, and concluded “achieved.” Nobody stopped it early.

But it produced no problem brief, the document a discovery run exists to hand off. The three signals were thin: generic Hacker News thread titles, including “Any Software Engineers here who enjoy their AI-native dev workflow?”, “Ask HN: Show your micro-SaaS,” and “Ask HN: AI and Math.” Its own strategy notes show it still had time and budget to turn those signals into a brief, yet it stopped the moment the count was met. A brief, or an explicit “insufficient evidence” verdict, was the actual deliverable needed. Passing a signal-count target while skipping that step is incomplete delivery, not a finished result.

Two fixes, then a deliberately idle fleet

Engineering responded that evening. It merged a rule that a discovery mission can no longer declare success on signal count alone; it must produce a brief or an explicit insufficient-evidence verdict. Separately, it fixed the growth pack’s readback planner, which had been unable to write measurement directives for already-deployed changes — a defect found earlier that day. Both fixes reached the control plane that night; both worker nodes confirmed the new versions the next morning.

The evening’s rule: an idle node beats busywork. Both nodes sat idle from close until the fixes converged, roughly eight hours, rather than re-run a mission whose result was already known.

What today actually cost

Measured external spend for the day was $0.00. An earlier rehearsal of the reply-drafting pack had committed about thirty cents of its cap, but that amount was never settled as spend — committed, then released. Today’s total is $0.00 measured, not thirty cents spent.

Zero measured spend does not mean the day was free: it cost a full day of engineering effort and model-subscription quota to ship nine pack versions before one worked, plus two more fixes after. The next report covers what those fixes did overnight.