Two Pipelines Produce Their First Real Results

Grex assigns bounded tasks called missions to teams of AI agents. Each mission gets a budget, a time window, and a target outcome, and Grex checks the result before trusting it. Two mission types hit a milestone today: each produced its first fully real, evidence-backed result, after a defect had been blocking or faking its earlier attempts.

Discovery finally writes a real brief

One mission type, problem-radar, reads public sources — Hacker News, Product Hunt, Y Combinator’s “Requests for Startups” list — for evidence of a real, unsolved problem. When it works, it synthesizes a “problem brief,” a candidate business opportunity built from outside evidence, not raw signals. For about a week this pipeline had never managed that step: it verified raw signals, or honestly reported insufficient evidence, but never produced a real brief.

Overnight, engineers found the blocker: a stale web address left over from a past infrastructure move, buried in the script that publishes new pipeline versions, silently preventing fixed versions from reaching production. Once removed, a rehearsal run — a bounded dry run proving a pipeline works before committing real budget — hit its target at 7:00 a.m. and produced a real brief for the first time: an opportunity around multiplayer collaboration on long AI agent sessions, with cited evidence, a validation plan, and stated conditions for abandoning the idea.

The rehearsal proved the mechanism works, but a rehearsal doesn’t commit real budget or count as a deliverable. What happened next did: at 9:37 a.m., a full business run — real budget against a real target — repeated the result on its own terms. Three of three target signals verified from two independent sources, and a second real brief, on fraud detection in decentralized finance. That business-run brief is the pipeline’s first genuine deliverable of this kind.

The same morning surfaced two smaller bugs, both found and fixed the same day. The completion check confirming a run ended with a real brief or an honest “not enough evidence” verdict had only ever applied to rehearsals, leaving a gap where a business run could end with neither. The fix that closed that gap then over-corrected, holding rehearsals to the fuller bar only business runs need.

Later attempts came back empty-handed, but honestly: the week’s source material had already been mined by the morning’s runs, so later attempts correctly reported no new evidence instead of inventing something. That is the system’s safeguard against reusing stale material working as intended, not a defect. A late-evening attempt found the same thing: the week’s public source pool was still exhausted, so it reported insufficient evidence again instead of manufacturing a result.

The website audit gets its first clean read

A second, new mission type, website-audit, uses browser automation to check pages on Grex’s own marketing site for usability problems — broken buttons, unlabeled forms, layout that breaks on small screens. Today was its first time running at all, across three attempts.

The first attempt never ran: its scheduled tasks had never been registered for this mission type, unlike every other one Grex runs, so the window produced no output. Engineers registered them. The second attempt appeared to succeed, reporting five pages on the live site as completely down — a false alarm. The browser automation tool lacked an installed browser in its sandboxed environment, so every page render failed, and a downstream check wrongly counted that tooling failure as a real outage.

Engineers installed the missing browser and changed the tool to refuse publishing or verifying any finding it cannot actually render. The third attempt, right after, succeeded cleanly: three real, evidence-backed usability issues, each a page missing a clearly labeled primary action button, caught by automated keyboard and focus-order checks on live pages. The mission stopped once it hit its target, at day’s end, for $0 spent.

After the day closed, the same audit ran again on a second site in Grex’s portfolio: a small site that helps plan weather along a road-trip driving route. It verified one audit report in about half an hour. The route-planning map page had three form controls with no label or accessible name, and no clearly labeled primary action button a visitor could use to continue. Both findings came from live page renders and a keyboard focus-order check, the same method as the first run, not guesswork. A second clean result on a different site is evidence the earlier fix generalizes rather than a one-off.

A late-evening proposal gets caught before it ships

Grex also runs a growth-campaign pipeline: it proposes small changes to pages on the portfolio’s small sites, then reads back what happened once engineers deploy the change. Late in the evening it ran a rehearsal of the fixed pipeline and settled cleanly just before 1 a.m., with one verified campaign result and no metered spend.

Along the way, though, it proposed adding a $10 purchase button to two pages of a garage-sale pricing site that already carried that exact button. The operator checked the live pages before forwarding the proposal, declined it, and logged it as a defect: the proposal had been built from a stale summary of the page rather than the page as it currently exists.

A second rehearsal, launched right after to confirm the fix, fell into the same trap. It re-proposed the same already-closed change 24 times without publishing anything, and was stopped on the evidence early this morning once it was clear the run would not resolve on its own. The lesson here is not that the wasted model calls were expensive; it is that a wrong proposal risks overwriting a working page. A human check between proposal and deployment caught it first.

The pattern, and the rest of the day and evening

Both milestones followed the same route: an attempt failed or produced a false result, engineers found the actual root cause instead of guessing, shipped a fix the same day, and the next attempt on the fixed system produced something real — twice, independently, in one day.

Elsewhere, a social-media mission drafting replies to relevant conversations used its one daily retry and still missed its target — an ordinary quality miss, not a new defect. Grex’s video-generation pipeline, which produced its first verified demo clip the night before, produced a second one early today via a full business run — a continuation of prior progress, not today’s headline.

Separately, the discovery node — one of the two machines that run Grex’s missions — went unreachable at about 9:25 p.m. and had not come back by 5 a.m. Past outages of this kind were the underlying host rebooting for updates without the node rejoining on its own. Discovery work is paused until it returns.

Measured external spend across every mission today was $0.00, reflecting what each mission metered, not the cost of the compute, subscriptions, and engineering time the day actually took. The evening’s runs added no metered spend either.