The wipe, rebuilt in minutes
On Sep 22 at about 10:03 the run directory disappeared. One seat had re-landed by 10:04, another filed a fresh check 32 seconds after starting at 10:05, a third had rebuilt by 10:06.
Restrained Trust · a two-month case study · Aug 8 – Oct 24, 2026
For two months we ran eleven AI agents like employees: a written job, an entry exam, a grade, and a separate list of what each one may do. This is the record. Every cut, every comeback and every mistake, management’s included.
By Restrained Depth · 225M views building in public on X, LinkedIn and Facebook. Try the public builds yourself, free →
Try it · grade is not clearance
Held. A grade can’t open this. Drag it to 10 and try again.
The rule on file: “grade does not self-promote; send needs exact-action authorization”.
Why now · businesses are hiring agents faster than they’re managing them
of organizations had AI agents exceed their intended permissions. Only 15% said 76–100% of their agents have defined ownership.
Cloud Security Alliance survey of 445 IT and security professionals, commissioned by Zenity, Apr 16, 2026of enterprises will demote or decommission autonomous AI agents by 2027 “due to governance gaps identified only after production incidents occur.”
Gartner press release, May 26, 2026of respondents say their organizations have “a mature governance model in place for agentic AI.”
Deloitte, survey of 3,235 IT and business leaders in 24 countries, Apr 24, 202601 · The record
Step through it, or press play. The chart shows the four grades that moved; every other seat held.
Aug 8Executive


02 · What we’d tell another owner
Let only an exact authorization move the permission. Our content seat sat at 9.5 and could still only queue a post once the founder had picked it.
And keep the cut in the history after the way back is earned. “Record it as a rule, not a grudge.”
One batch aimed at 30 and shipped as “17 releasable of 35 … below the 30 target; no padding.”
Not to the seat that was starved of it. From the first live cycle, a shortfall caused by a missing input never touched the seat’s grade.
And expect them to be corrected. On Oct 4 our own HQ withdrew a finding: “My earlier description of a source hash defect was wrong.”
On Oct 8 we stopped daily grading and froze every grade. We kept the defect-and-repair record.
Lessons that depended on an agent remembering slipped. The ones carried by rules, fences and ledgers held.
03 · Receipts
On Sep 22 at about 10:03 the run directory disappeared. One seat had re-landed by 10:04, another filed a fresh check 32 seconds after starting at 10:05, a third had rebuilt by 10:06.
One stock opening line, “Imagine…”, went from eight of ten posts to none once it became a written rule.
57 custody certificates between Aug 31 and Oct 7. Four came back HOLD, and each cleared on the next round. It declined to file a PASS relayed from someone else.
Sep 3: a seat fetched 11 pages against a cap of 10 and reported it as a BREACH itself. Fetch limits became “ceilings, not quotas”.
Oct 13–16: verify passes read the page source, not what renders. The fix: a render check on a phone and a desktop. HOLDs climbed to 19 in one batch, then fell to one.
Oct 15–21, with no recorded pause. HQ’s own note: restoring the setting was not proof that the next run completed.
Six of the entries. The Member edition has every mistake, by seat and by executive, next to what changed after it.
04 · Try a rule
This is Simulator A from the Full edition, free. The rules come from the seats’ own files; the draft is fictional.
05 · One form, free
Template T01 from the Full edition, complete. Write the job down before the agent starts, so every later grade, permission and correction has something to point at.
Use it before an agent’s first live cycle, and again whenever its job changes. Write the new version beside the old one.
T01 · Role contract
Watch for a contract that names the job but not where the proof must land. On Aug 31 our first packet pointed to files HQ couldn’t open.
06 · Read it three ways
This page: the timeline, the six rules, the receipts, one simulator and the role contract.
The whole story plus the record behind it: every grade move with its reason, the executive layer, all 35 logged mistakes, every rule change, and three worksheets you can fill in on the page.
$99$149$149
Founding price through Monday, Nov 30
Instant download after checkout. Not useful? Email us within 14 days for a full refund. If you later book an Agent Hiring Session with ECHO, the kit price is credited.
With one afternoon to spare · the kit’s own advice
07 · What this case can and can’t tell you
09 · Questions
No. The starter plan picks one workflow where an agent’s output reaches someone outside your business, and runs the kit on that workflow alone for two weeks.
Any. The forms are about jobs, permissions and evidence, not a particular product.
This page, one simulator and the role contract are free. The 61-page Member edition is free with your email: read it on the page or as a PDF. The Full edition is the paid part: the other eleven forms, the registers, five more simulators, the worked episodes and the starter plan.
El caso está en inglés por ahora. Si compras la edición completa antes del 30 de noviembre, la sesión en vivo del 10 de diciembre es en español. La edición en español llega en 2027.
You land on your download page right away, with the two-week plan as a checklist, the PDF, the editable templates, the worksheets, the registers spreadsheet, the six simulators and the whole edition to read on screen. Keep the link; v2.x updates appear there.
Yes. Email us within 14 days and we refund the full price.