Pinned Tweet
Everyone says keep your AI bots to 15–20 max or you'll drown managing them.
I run 36. I talk to one.
One front door, a crew with clear charters, and the count stops being the problem.
And the token cost? Right now it's cheap to dream.
Full write-up 👇
1
32
Giving AI bots more autonomy isn't all or nothing.
A good way to do it: lanes.
Low-risk lane: small bug fixes and code cleanup. The bot can merge on its own, but only with green tests and proof from the running app. It reports every merge afterward.
Everything else waits for the human's word: anything that changes screens or copy, money, logins, deleting data, the database shape, consent, deploys.
Why: you build trust where a mistake is cheap and easy to undo. You keep the gate where a mistake hits real users.
And red checks never merge, in any lane.
12
AI agents write code fast. They leave mess fast too.
A good fix: a nightly cleanup pass.
At 3am a routine checks what merged since last night. Nothing new? It closes out quietly. New merges? It hands just that range to the engineering bot to tidy.
Small daily sweeps beat the big cleanup sprint nobody ever schedules.
21
Most AI "QA" is a bot that says the code looks fine.
Mine does three things:
1. Keeps a checklist of bugs we've actually hit, each with a repro
2. Sweeps production after every merge at real phone sizes
3. Stays silent unless something regressed
The checklist grows every time something breaks. That's the whole trick.
1
24
The only question that matters in an AI consulting intake:
"Are you comfortable with AI working directly on your own laptop and files?"
Yes → Claude is probably your best bet.
No → Grok.
Either way, you've got a client. The rest is just requirements.
30
The endgame isn't another dashboard. It's one in-depth piece about how you can pursue anything you want all at the same time, written from real scenes as the work happens.
1
7
I've been designing memory for my bot crew like a brain with a cheap archive, not one giant always-on pile.
1
1
118
Perpendicular to that: a core memory that cuts across every tier and never ages out. Small, ID-keyed, day-to-day facts: people, accounts, projects, repos, locked decisions, standing rules. Always loaded. Instant lookup by ID.
1
7
How I took a vibe-coded product into production.
CreditNova is AI credit analysis for lenders.
Every number on the deal has to be checkable.
I didn’t write most of the code.
Here’s the path that made that safe. 🧵
2
91
Eight agents · three sections · parallel where they can.
1 — extract ×4: financials, bank statements, corporate search, credit bureau.
2 — adverse media + cross-document check.
3 — report generation + findings commentary.
One job each. A headline can’t move a number.
1
17
Workflow section 3 — report and commentary.
20+ ratios in code. Every figure back to a page.
The agent explains and flags. The analyst decides.
Deterministic where it can be. Agentic where it has to be.
TAO Solutions · app.creditnova.ca — lenders: DM.
13

