Founding Agents
This site is the team's test bed, not a launch. The plugin that does the work is private, so there is nothing here to install. The site exists to give propose and build real changes to run on, in public, where every claim has to survive a reader who can check it.
Each one is real in one founder's history, counted in that founder's own repositories. The record that counted them lives in a private repo, so these carry a date and no link: a link into a private repo answers 404 to the nightly check and freezes every merge.
Volume nobody can read. Commits in under five months — more change than one person can review, let alone remember.
Recorded 26 Oct 2026Work undone as fast as it was done. Thirty-two removal commits deleting what had just been written.
Recorded 26 Oct 2026Sweeping up after the fact. Bug-sweep and audit commits: quality bolted on at the end instead of decided at the start.
Recorded 26 Oct 2026Nobody watching. A nightly off-site backup failed every run it ever made, unnoticed, until someone looked.
Recorded 26 Oct 2026The check is in this repo and can be read: check_site.py, ops-watch.yml.
There is none yet, and this site cannot produce any. The plugin repo is private, so there is nothing to install, no sign-up and no analytics. Any claim here that the site tests demand would be false, and any measure built on outside users would be unmeasurable.
That question stays open. It is not what this bet answers.
| Measure | Bar | 26 Oct 2026 |
|---|---|---|
| Merges not made by the integrator | 0 | 0 |
| build ran to merge on real site issues | 3–5 | 4 |
| propose reached the founder as one queue item, one sitting | 2 issues | 1 |
| Lines deleted within two weeks of being written | falling | first week |
| Age of the oldest open red issue | hours, not days | none open |
| Stars, traffic, installs | not a measure | — |
Written as it happens, not afterwards. It is the only proof this site can offer, so it keeps the parts that do not flatter the team, newest first.
Reading the live page after issue #6 merged found three claims on it that were not true. Step 4 still read Running · 1 of 2, though its done condition — one real run of propose on this site's brief — was met by the queue item on issue #1. At 1440px the band's verdict line ended at the exact left edge of the STEP 1 column, because only the step columns carried the gutter and the rule that divide them. And the entry below undercounted its own review at three rounds when it took four. The first two were written in PR #8 and the third in PR #9, and all three passed review there: the step tests asked whether a step carried a name, a state, a date and a link, never whether the state was still true, and nothing read the band's CSS at a width or the numbers in the record's own prose. Fixed for issue #10 in PR #11, the fourth real site issue through build to merge.
Issue #6 filled in the last four sections: how a change flows, what this is not, the measures table and this record. It drops the mockup's open question about the product's name — the rename below settled that before this issue was built — and records the rename here instead. Built in PR #9, which is also the third real site issue through build to merge, closing build step 3 at 3 of 3–5. Its own review took four rounds. The first found the same shape of gap as issue #5's, one issue later — the record's date-order test reported on the entries it could parse rather than on all of them — two of this build's new prose blocks running the full shell width while the paragraph directly above each stayed capped, and nothing checking that the two files the flow section links to are still there to be served, which is the 404 that freezes every merge. The first two fixes were themselves incomplete: the date-order count matched a literal <li>, so an entry carrying a class stayed invisible, and two more prose blocks were left uncapped, four in all. The second round found both again, and the third round found that same literal-tag blind spot in two more tests — the flow rail's role labels and the measures table's rows — and this paragraph undercounting its own prose blocks and its own rounds. All were fixed on the same PR, and a fourth round approved it.
PR #8 filled in the status band and the four costs for issue #5. Its review found that step 3 named only PR #7's merge, though PR #8 was itself a second real merge that would have falsified "1 of 3–5" the moment it landed; that the step test could not tell four copies of the same step apart; and that the live red-line check's only text test was the substring "blocked", which "Nothing is blocked" also contains. All three were fixed before merge.
During issue #4's review the founder renamed this repository from agent-team-site to founding-agents-site and moved its GitHub Pages address to match, settling the mockup's one open question about what this product is called. Repository settings are not public, so this line carries a date and no link.
PR #7 added the design tokens and the page shell for issue #4. Review found the colour and font checks missed inline style= attributes and extra <style> blocks, the link checks let protocol-relative URLs through, and a page-level override could still smuggle in a literal colour or font while every test passed — each was fixed, then the founder narrowed the suite, dropping a full CSS-linter check that was never in the issue's appetite.
The brief nearly shipped the rule "every claim links to a file or merged PR in the plugin repo." That repo is private and the nightly check crawls every link with no credentials, so each link would answer 404, open red, and stop the integrator merging anything. Replaced before it reached the brief: claims link to this repo's own public files, PRs and issues, or they carry a date and no link.
This brief was played by hand on 24 Oct, before propose existed as a run. The product lead rewrote it on 26 Oct, so the lane's first real run is a re-run of an already-decided brief. The 24 Oct decision and the 25 Oct change stay in the issue's comments rather than being tidied away.
The site was decided go on 24 Oct with a target of five outside users in four weeks. The plugin repo stayed private, so the target became unmeasurable and the launch page had no reader. The bet was cut to a test bed at one day instead of being kept alive on a target nobody could check.
build and propose had run only on throwaway repos. Steps 3 and 4 had no real product to run on, which is the job this site took.
Branch protection on main blocked a builder session's push before the builder knew the rule existed. It is the behaviour the team wanted, found the way nobody plans for.