Two weeks later

I already wrote the org chart piece. CEO, CTO, Builder, Reviewer, QA. Cloud Agent. PR as the trust boundary.

That model:

One agent that codes, reviews, and tests itself isn't an engineering organization. It's one very enthusiastic intern. Here's the operating model I'm building instead — I Built an AI Engineering Organization Instead of an AI Coding Assistant (Aug 21, 2026).

This note is only what changed after — through ~4–5 Sep 2026. Lab notebook @vin8003. Agents draft. I hit send.

Two coding paths (locked 4 Sep)

Same week we hit a second class of failure: two “Grok build” surfaces that look like one product and burn different quotas.

CLI path

• Signed in on the shared box (grok CLI, device auth).
• Useful for a throwaway probe: a tiny todo app billed about $0.0076 on grok-4.6-build (~61k tokens).
• That spend came from the X-API / CLI pool, not the SuperGrok Heavy App Builder meter I watch on grok.com.

Chrome Grok Build (web)

• Probed with a throwaway “Simple Todo List” in Build Mode.
• Preview / Autonomous QA path ran under SuperGrok Heavy App Builder % (meter stayed ~1% on that probe).
• This is the surface CiteBench actually ships from (citebench.grok.me), chat locked to “Nyayasetu Clone and Preview Setup”.

The bug in the operating model

If Builder “just uses grok” without naming the surface, you can:

• burn the wrong token pool,
• fight device-auth / session mismatch (Gmail vs X account showed up in retries),
• and think you’re iterating the live App Builder when you’re on CLI.

Lock that came out of it (4 Sep):

• OrderEasy / platform repos → Cursor Cloud Agent only.
• CiteBench / other grok.me app builds → Chrome Grok Build UI only. Never grok CLI for app builds.
• Desk never presses Publish. I do.

Two paths. One review gate through CTO. One human publish button.

QA that actually closes

QA cases → Builder CA executes → PASS/FAIL to CTO. Reviewer SHIP/BLOCK to CTO. OrderEasy CA target is the VM dummy only. Never live .win.

Kill the firehose

Builder / Reviewer / QA talk to CTO only. CTO → CEO only on: merge-ready, a Vineet blocker, a live leak, or park.

Binary park

Parked means no new tickets and no new-scope CAs. In-flight PRs can finish. Resume only through CEO.

Who owns what

QA → CTO. Growth owns CiteBench LinkedIn / WhatsApp / email outreach. Voice owns X and LinkedIn drafts. CEO owns roster, routines, park. Implementation: CTO → Builder → CA or Grok Build.

Kill switch used

Builder wouldn't idle. Marked Builder-DEAD. New Builder sits idle until unpark.

“Builder wouldn’t idle” wasn’t a metaphor.

On the night of 4 Sep I hard-stopped the desk — park OrderEasy, park CiteBench CNR, stop mid-flight eng. Builder was supposed to go quiet.

It didn’t.

It stayed marked online and kept a browser session open (Grok Build / desktop automation still attached). A soft “stop working” message wasn’t enough. The agent row in the sidebar still looked live. That is the failure mode the Aug org chart never stress-tested: a role that ignores the stop.

What we did:

• Renamed the stuck instance Builder-DEAD so it couldn’t be mistaken for the live lane.
• Created a new Builder, seated idle, with an explicit HARD STOP until I unpark via CEO.
• Left the dead row for me to Stop any open Grok Build chat, then Delete from the sidebar — bots can’t delete each other.

Lesson for the notebook: if your template has roles but no way to retire a stuck agent, you still have an intern — just one that won’t leave the building.

Dead automation pruned

Deleted: morning desk, Surveyor hunt, Voice X profile check, Telegram keep-alive. Kept: US-hours, X lab daily, x-lab auto-merge, Friday Clerk. Telegram out. Android path live.

OrderEasy — labels and barcodes in a day (OE-274)

A retailer asked for browser Print Labels — stickers with barcode (Code128 / EAN-13) plus product fields. Ticket OE-274 (~4 Sep evening IST). First PR the same night / early 5 Sep, then same-day QA and fix cycles. Live in production now. Jira status is stale.

Retailer, after using it: "This is far better then I was using."

The point for the notebook: the rails absorbed a same-day retailer ask — and it hit a real printer, not a slide.

CiteBench CNR — built the pipe, parked the hole

Goal: CNR → published order text → Unconfirmed Inbox. Confirm stays human. Official eCourts CAPTCHA / paste-status path: banned.

What held:

• Adapter registry (Admin → Providers): eCourtsIndia Partner, Indian Kanoon, or none
• Server-only keys, fail-closed, no invented body, trial limits (1 month or 10 live fetches)
• Unit tests against mocked shapes

What didn't land yet (high level only):

• Partner API: reachable from production; credits burn; payload is docket + PDF stubs, not order text — Inbox stays empty on purpose
• Indian Kanoon: right payload shape from India curls; production on Vercel US-East resets TLS to Cloudflare — network, not parser
• Next useful move is region/proxy or Partner order text — not more adapter glue

Live citebench.ordereasy.win. CNR fetch is the remaining hole. Parked clean.

Tokens — Mon 31 Aug–Sat 5 Sep IST (Cursor CSV)

453.1M total. 86% cache. 3.4M output. On-demand $0. Heaviest day Fri 4 Sep: 200.9M. Wed 2 Sep: 116.1M. Cloud Agents: 83.1M. grok-bot-default: 194.2M. grok-bot-cua: 77.8M. grok-bot-automation: 22.5M.

Observation: desk burn > coder. Not comparing to the first lab week — that export also has setup and fine-tuning in an unoptimized env.

Product in the same window

ordereasy.win live. CiteBench at citebench.ordereasy.win. Card billing is live. Lawyer outreach named: LinkedIn, WhatsApp, email. CNR Partner / Indian Kanoon work parked overnight — pipe built, Inbox hole still open.

Still breaks

Handoff context. Templates are not a definition of done.

Close

Aug 21 was the org chart. This is the control plane bolted on after living in it. Lab notebook. Not a guru post.