Agentic maturity · Rerun November 2, 2026

Eight rungs done. Nine and ten are closer, and not done yet.

On July 20, 2026 we graded our own system against the ten rung agentic maturity ladder from Search Atlas. On November 2, 2026 we ran the same audit again, against what runs today. Every verdict below carries a date and a source you can check. Where a rung got weaker, we say so.

0 of 10

rungs done, same as July 20

0 work commits since July 20 0 desk jobs since October 4 0 live desks First run July 20, 2026

What changed since July 20

The system was rebuilt around ten desks and a queue.

2026-09-25Hermes became the main system. Claude Code went read only as the way back. (commit b50296df88)
2026-09-25Ten roles became ten live agents, each with its own profile and a SOUL file built from the org chart. Ten SOUL files are on disk today.
2026-10-04A job queue runs desk work in its own branch, away from any chat. Since October 5 every finished job gets an automatic review. 3,252 jobs and 3,084 reviews so far.
2026-10-20The driver seat. Deploys and internal work run without a click and are logged. Justin keeps three stops: client sends, prices and spend. 2,337 logged actions since October 13.
2026-10-14A blog release engine started publishing client posts on a weekday schedule, with no one starting it. Four sites release posts this way today.
2026-10-26Justin and four teammates can message Marshall, the lead desk, in Slack. An iPhone app signs in to the system over a private network.
2026-10-29Setback. Two client blog releases failed for days and no alert reached anyone. Network of Directors missed 3 runs and Celebrate staging missed 5. Network of Directors was fixed the same day.
2026-10-01Narrowed on purpose. Error fixes now wait for Justin's yes in Slack before they ship. The self heal loop made 3 fixes since July 20, and 1 shipped on its own.

The scorecard

Ten rungs, July against today.

Rung	July 20	Nov 2	One line of evidence
1 · Real work in version control	DONE	DONE	20,858 commits on master; every desk job works in its own branch
2 · The agent in your terminal	DONE	DONE	HERMES.md is the rulebook since 2026-09-25
3 · Fully agentic	DONE	DONE	27 tool servers on one desk today, 21 of them client sites
4 · Orchestration	DONE	DONE	Ten desks and a job queue, 3,252 jobs since 2026-10-04
5 · Always on	DONE	DONE	20 scheduled tasks registered today; a weak spot is named below
6 · Spec driven and self checking	DONE	DONE	232 goal lock verdicts since July; 1,692 desk jobs sent back by review
7 · Multi model council	DONE	DONE, weaker	254 councils since July, none with a second model family
8 · End to end and remembering	DONE	DONE	971 memory notes; the nightly insight run ended clean on 2026-11-01
9 · Days at a time, unattended, on rails	WORKING	WORKING	Client posts ship on schedule; the alarm failed for days in October
10 · Commander level, from your phone	WORKING	WORKING	Slack reaches ten desks; one clean rolled up briefing is still missing

The climb · Rungs 1 to 10

Every rung, dated and sourced.

Dates and counts were read on November 2, 2026 from git, the decision log, memory notes, Task Scheduler and live web checks. July's evidence still stands unless a line here says otherwise.

Real work, in version control

WAITING

2026-04-02First commit. Every change since lives in git and can be undone.
2026-11-0220,858 commits on master. Since July 20: 5,901 work commits plus 11,355 automatic backups.
2026-10-04Every queued desk job works in its own branch and worktree. This rerun ran in one.

For your business: every change is tracked and can be undone.

The agent in your terminal

WAITING

2026-09-25HERMES.md replaced CLAUDE.md as the written rulebook every desk reads first.
2026-10-04A house rule went into all ten SOUL files: no inline code in unattended runs.

For your business: the system follows a written playbook, not a guess.

Fully agentic

WAITING

2026-11-0227 tool servers wired to one desk: Search Atlas, GitHub, PostHog, design and files, plus 21 client WordPress sites.
2026-10-02Desks call Search Atlas directly with their own key, no person in the loop.

For your business: it works inside the tools that get you found.

Orchestration

WAITING

2026-09-25Ten roles became ten live desks: lead, inbox, memory, marketing, revenue, clients, builds, review, skills and safety.
2026-10-04The job queue went live. 3,252 jobs have run through it.
2026-10-29Up to 12 jobs at once, and up to 6 helpers inside each job.

For your business: many jobs at once, each checked before it goes out.

Always on

WAITING

2026-11-0220 scheduled tasks are registered: blog releases, review feeds, backups, Search Console watch, nightly insight and the job queue.
2026-10-21Hosting moved to Vercel. July's line about Netlify hosting is retired.

Weak spot

The four release tasks I read run only while the desktop is signed in. The October 29 rule says they must run signed out. Read from Task Scheduler on November 2.

For your business: posts and checks run while you are on a job.

Spec driven and self checking

WAITING

2026-11-02296 planned runs since July 20. 232 of them carry a goal lock verdict.
2026-10-05Every finished desk job gets a review. 1,356 passed and 1,692 were sent back.
2026-10-20Every deploy folder is screened for key files before anything uploads.

For your business: nothing goes out until it passes its own check.

Multi model council

WAITING

2026-11-02254 planned runs since July 20 went through a council of advisors before the plan was locked.
2026-04-13July's proof was a review across Gemini and Codex. That evidence is from April.

Got weaker

None of the 254 councils since July used a second model family. Twenty name Gemini or Codex, all as a topic. The cross family advisor is still listed as deferred. Several opinions, one family of models.

For your business: big calls get more than one opinion before we act.

End to end and remembering

WAITING

2026-10-13The July instinct pipeline was retired. Its 54 lessons were sorted into memory notes.
2026-11-01The nightly insight run ended clean at 9 PM.
2026-11-02971 memory notes. Every desk searches them before it touches earlier work.

For your business: it learns from each job and keeps the lesson.

Days at a time, unattended, on rails

WAITING

2026-10-14Client blog posts ship on a weekday schedule for Bergheim Plumbing, Network of Directors, Celebrate Rentals and our own site. The guard refuses any change but the posts that are due.
2026-11-02Bergheim's post went live at 6:04 AM from a 6:00 AM task, with no prompt. Receipt below.
2026-10-06When a job passes review, its next step runs on its own. 80 follow ups, 75 done.

Still a gap

From October 21 to 29, two releases failed for days and no alert reached anyone. On November 2, two of the four release tasks I read ended with an error code. The engine has run 19 days, not the 60 to 90 days of proof the roadmap asks for.

For your business: posts go live on your schedule while nobody is at the keyboard.

Commander level, from your phone

WAITING

2026-10-26Justin and four teammates message Marshall in Slack, and Marshall hands work to the ten desks.
2026-10-21One Slack line starts a monthly webinar pack that runs across four desks.
2026-10-13Justin answers the three stops in one word from his phone.
2026-10-26An iPhone app signs in to the system over a private network. Five sign ins were logged that day.

Still a gap

There is no single rolled up briefing yet. Our agents posted 803 Slack messages in 7 days, and 61 percent were job notices. Justin switched six kinds off on November 1. The digest that replaces them has not started.

For your business: one day you run the whole operation from your phone, between jobs.

The receipt · Real, not a sample

What one unattended run left behind.

In July this spot held a sample with made up values. This one is real. It comes from the release engine's own record of its November 2 run, read back from the live site the same morning.

Rung 9 receipt · Bergheim Plumbing blogReal run

Started6:00 AMLive6:04 AMFailures in a row0Read back200

Job

Scheduled task Blog Release bergheim-plumbing, weekdays at 6:00 AM Central

What it shipped

One blog post: main-water-shutoff

What it may touch

Only the blog posts whose date has come. The guard refuses anything else.

Who started it

Task Scheduler. No person and no prompt.

Same morning, two misses

Celebrate's 7:00 AM task and our own 6:00 AM task ended with an error code. Both are flagged for their owners in the audit notes.

What this receipt does not prove

Days of runs with a working alarm. That is the rung 9 gap above.

Phase A planned this receipt from the error self heal loop. That receipt was never built. The loop made 3 fixes since July 20, so the release engine is the better proof.

The roadmap · What flips 9 and 10

Three phases, where each one stands.

Phase A · Prove it

Done, by a different loop.

A real receipt is on this page. It comes from the blog release engine, not the self heal loop the plan named.

Phase B · Clients run on rails

Working since October 14.

To call rung 9 done: every release alerts on its first failure and runs signed out. Then about 60 clean weekdays, each with a receipt.

Phase C · Run it from a phone

Working, the hand off is real.

To call rung 10 done: one rolled up briefing replaces the job notices. Then one dated case where a phone message runs across desks and comes back once.

Why we publish the misses

An owner who got burned before checks the receipts.

Our clients have heard big promises from agencies before. So this page shows the failures next to the wins. A system that grades itself honestly is one you can check.

Rungs done, dated and sourced

Rungs working, with the gaps named

Rung that got weaker since July

Scheduled tasks registered today

Self check · Computed in your browser

This page checks itself.

The line below is counted when the page loads. Nothing in it is typed by hand. If a count is off, it prints FAILED.

SELF CHECK· 0 done· 0 working· 0 sourced lines· 0 em dashes· 0 console errors· CHECKING