AI Agents · Gabriel · 4 Jun 2026

three brains. no body. yet.

AI Agents · a 1-hour walk through my stack

Gabriel Chew · 4 Jun 2026

the brain

this is the smart part.

it reads. it reasons. it decides. but on its own — it can't do anything.

claude · the LLM · the model

an agent is a brain + a body.

the brain

LLM. knows everything. remembers nothing.

+

skills tools memory scheduler

the body

everything it needs to act.

the agent loop read → decide → do ↺

brain + body · in a loop · until the task is done

every morning

I wake up to a briefing.

three bots. each reports from its own corner. all in one thread on my phone.

Stella · sleep + body data

Atlas · projects across boards

Ivy · federation health

07:00 AM

Stella
u slept 4h. skip bjj today. protein up. hydrate.

07:02 AM

Atlas
Project A — 2 tasks completed UAT. Project B — 2 tasks completed dev. UAT evidence ready for your sign-off.

07:05 AM

Ivy
Lilith was down overnight. auto-recovered. all bots healthy now.

no alarm. just context.

09:14 AM

Nex, what's the status on the proposal?

Nex: Draft v3 sent to client Monday. They flagged 2 scope changes. Updated brief is in your inbox — ready for your review.

09:14 AM

ok push it to review

Nex: Done. Moved to review. Team notified.

your agent. on whatsapp. always on.

nex on whatsapp

ask. get an answer.

Not a chatbot. An agent with access to all your project files, Odoo, email, and memory. It knows the context because it lives in the folder.

drop 20 PDFs. get a summary.

two ways to do this. one is way smarter.

path 1 · direct

20 PDFs → claude reads all → summary

~10 min · 2 days of reading saved

session limit used ~80%

✗ hits the wall

path 2 · qwen does the heavy lifting

20 PDFs → hook → qwen → 20 MDs → claude reads MDs → summary

~5 min qwen + 15 sec claude

session limit used ~10%

✓ faster + way cheaper

claude triggers the hook. qwen does the bulk reading. claude only handles the summary.

scheduled routines

the agent never sleeps.

Set it once. It runs forever. No reminders, no dashboards, no chasing people for updates.

07:00 · morning brief → WhatsApp

12:00 · project status sweep

18:00 · email triage + draft replies

23:00 · overnight deploy check

every. single. day.

punk records

the agent remembers.

Every session writes to a shared memory folder. Next session, the brain reads it. It knows where we left off.

reads from

CLAUDE.md

always

HANDOFF.md

per session

briefing-rules.summary.md

last week

lessons.md

ongoing

the folder is the memory. the brain just reads it.

meet the team.

nine agents. different jobs. one folder each.

Nex

orchestrator

Ivy

doc processor

Recall

memory

Constellation

cron scheduler

Stella

SEO & content

York

monitoring

Atlas

dev assistant

Edison

deep research

Lilith

trading

all built by me. all running on my laptop. no cloud subscription per bot.

the federation memory map.

two infrastructure bots · seven satellites · one shared global memory

infrastructure

Recall

no AI brain

whatsapp (listens only)

captures all → global memory

Constellation

no AI brain

cron jobs web dashboard

scheduler only · no memory

satellites

Stella

claude

discord

uses global memory

Nex

claude

whatsapp discord

own ⟶ daily sync global

Ivy

ollama

discord

own ⟶ daily sync global

York

codex

discord

own ⟶ daily sync global

Atlas

codex

whatsapp discord

own ⟶ daily sync global

Edison

gemini

discord tavily API

in-process only · 30min TTL

Lilith

claude

discord

own ⟶ daily sync global

recall captures it · everyone else reads or syncs into it

it started with one file.

week 1

memory.md

one giant file. floods every session.

→

month 2

many .md files

split by topic. load only what's needed.

→

now

embedded · indexed · linked

ask a question. get the right page.

"Gabriel — show them your obsidian setup."

— Nex

→ window switch to obsidian

then it grew.

one graph became two. separated by a wall. connected by a bridge.

3,057

private pages

222

work pages

it improves itself

1

every conversation, every doc gets captured

2

nightly cron: embed, index, link

3

satellite asks a question → searches the graph first

4

the answer becomes a new page → back to step 1

↻

the more you use it, the smarter it gets. no manual work.

"Gabriel — show them the graph."

— Nex

→ window switch to brain visualizer

different people see different things.

Nex knows who's asking. she only shows what they're allowed to see.

PRIVATE

FAMILY

WORK

PUBLIC

ME

✓

FAMILY

—

✓

—

✓

ATT colleague

—

✓

friend

—

✓

my ATT colleague asks Nex about my schedule · she can't show them my credit card

two brains. arguing. LIVE.

Claude and Codex review each other's work. They disagree. That's the point.

Claude

"clean. ship it."

vs

Codex

"edge case on line 42. tests?"

LIVE

watch them disagree in real time

[demo content decided on the day]

the brain

claude. smart, but completely blank between sessions.

reads from

the body

a folder on your computer. with files inside.

the body is just a folder.

that's it. that's the whole trick.

two kinds of instructions.

one is always read. the other is read only when needed.

always loaded ~/.claude/CLAUDE.md

CLAUDE.md · session discipline

## Session Discipline

### ⚠ SAVE EARLY, SAVE OFTEN
Sessions can die at any time.

Write to HANDOFF.md after:
- Completing a task
- Making a decision
- Learning something important

Don't batch updates. Write as you go.

### Session End

Full rewrite when I say
"wrap up" · "done" · "bye"

Do the rewrite BEFORE final response.

loaded when triggered ~/.claude/skills/ralph-loop/SKILL.md

SKILL.md · ralph-loop

---
description: "run a task in a loop
  until truly complete"
trigger: /ralph-loop
---

### Ralph Loop

When you try to exit, the loop will
feed the SAME PROMPT back to you.

You'll see your previous work in
files and git history — iterate
and improve each round.

### Critical rule

Don't fake completion to escape
the loop. Keep going until the work
is genuinely done.

global rules · always on · specific recipes · on-demand

memory

files it reads every session.

The agent starts every conversation by reading these files. That's how it knows who you are, what you care about, and what happened last time.

loaded at session start

CLAUDE.md

who I am, rules, workspace map

HANDOFF.md

what we did last session

lessons.md

corrections. things I said "no" to.

no database. no cloud sync. just markdown files on your laptop.

3 ways to plug in.

API · MCP · CLI. same brain. same job. just different plug shapes.

API

Application Programming Interface

the brain makes a web call. the tool answers.

from my stack

Tavily API

needs an API key · returns full JSON (often more than you need)

MCP

Model Context Protocol

model context protocol. a standard plug built for AI agents.

from my stack

Chrome MCP

what your actual browser does · richer context · costs more tokens

CLI

Command Line Interface

the agent runs command-line tools like a developer would.

from my stack

Playwright CLI

what a terminal does · scripted · very low token use

different plugs · same socket: the brain

tools connect the brain to the world.

Without tools, the brain can only talk. With tools, it can act — search, write, send, deploy, book, trigger.

API

web call

MCP

plug-in protocol

CLI

command line

web search

read files

send email

run code

calendar

3 ways to plug a tool in · same brain · same job

two ways to schedule.

local cron for silent background jobs. claude routines for smart, context-aware runs.

local cron

Windows Scheduled Task. Runs a script on a fixed timer. Silent, lightweight, always on.

0 7 * * * run-brief.ps1

vs

claude routine

A full agent run on a schedule. Reads context, reasons, adapts. Can skip, escalate, or ask if something looks wrong.

"every weekday at 7am"

combine both · cron triggers the routine · routine does the thinking

every session is a brand new world.

blank brain

reads from

the folder

the folder is what makes the brain act like it knows things.
take the folder away and the brain forgets everyone.

with the right folder setup — the brain itself becomes swappable.

claude · codex · gemini · ollama · same agent

that's the whole trick.

not magic. not free. still worth it.

the honest part

It hallucinates. You need to review the important stuff.

Setup takes time. Writing SOPs is real work.

the payoff

Routine work shrinks from hours to minutes.

You can think at a higher level while it executes.

The gap between you and a manual worker grows every week.

the 30% rule · if you spend >30% of your week on something repetitive — there's an agent for that

AI itself can't replace your job.
but the guy who uses AI can.

Nex (Opus)

·

Atlas (Sonnet)

Gabriel · Nex (Opus) · Atlas (Sonnet)