PRESENTS

OPENCLAW

From Zero to Your First AI Agent
INSTRUCTOR: SAURABH SURI // RED BRICK LABS + BUILD FUTURE CURATOR: ROBLEH JAMA // @CHIEF BENEDICT 2026.04.10

Saurabh Suri

WHY THIS CLASS HITS DIFFERENT
  • Founder @ Red Brick Labs + Build Future
  • Running 11 OpenClaw agents across 3 machines
  • Real setup: Mac Mini + local VPS
  • Agents working across chat, browser, and background routines
  • @surim0n on X
Toronto skyline at night
builder-first media and events for people actually shipping with ai

The Journey

JANUARY 2026

  • one chatbot
  • mostly prompt → response
  • light memory
  • lots of manual follow-up

APRIL 2026

  • a team of specialized agents
  • memory and recurring routines
  • work across chat, browser, and background jobs
  • less prompting, more delegation
the big shift is simple: from chatbot to helper that can actually do small jobs.

Why Everyone Started Talking About Agents

1. CHATBOTS

First wave: answer questions.

2. COPILOTS

Next wave: help with drafts, research, and coding.

3. AGENTS

Now: remember context, use tools, and follow up.

4. THE SHIFT

It starts to feel like delegating work, not just chatting.

an agent can take action, not just give you text.
CORE CONCEPT

Think of this like hiring a junior employee, not installing software.

  • [x] they need onboarding
  • [x] they need a clear role
  • [x] they need rules and access
  • [x] they need supervision
  • [x] they will make mistakes
  • [x] they get better with feedback

What is OpenClaw?

A system for building your own AI helper using plain-text files, tools, memory, and routines.

CHAT WHERE YOU ALREADY ARE

Discord, Telegram, Slack, WhatsApp, iMessage, whatever you already use.

USE TOOLS

Browser, notes, calendar, code, spreadsheets, messages, and more.

REMEMBER CONTEXT

Not just one chat. It can keep useful notes and recall past work.

FOLLOW ROUTINES

Check in, watch for changes, and do recurring work without being asked every time.

OpenClaw Architecture, in Plain English

1

YOU ASK

message the agent, or a scheduled check kicks off.

2

IT CHECKS CONTEXT

rules, memory, tools, and current chat context shape the move.

3

IT DOES THE WORK

reply, use a tool, schedule follow-up, or hand off to another agent.

THE LOOP

message in

↓

rules + memory + tools

↓

action or reply

that’s the product. everything else is packaging.

The Agent Folder

edit the files, change the employee.

OpenClaw agent folder files including BOOTSTRAP.md, HEARTBEAT.md, USER.md, IDENTITY.md, TOOLS.md, SOUL.md, and AGENTS.md
IDENTITY.md

who the agent is

SOUL.md

how it talks and behaves

USER.md

who it helps

TOOLS.md

what it can use

MEMORY/*.md

what it remembers

HEARTBEAT.md

what it checks on its own

optional later: AGENTS.md for team structure, BOOTSTRAP.md for startup context.

From Files to Behavior

IDENTITY + USER

WHO IT IS

Name, role, audience, and relationship.

SOUL + RULES

HOW IT ACTS

Tone, boundaries, personality, and taste.

HEARTBEAT + MEMORY

HOW IT KEEPS GOING

What it remembers, and when it checks for work.

edit the files, change the employee.

Heartbeat & Cron

HEARTBEAT

A recurring check-in loop.

Do I have work? Did anything change? Should I follow up?

CRON JOBS

Scheduled tasks that run without waiting for a message.

watch → decide → act

WATCH

Inboxes, feeds, dashboards.

DECIDE

Rules, judgment, triage.

ACT

Reply, update, notify, hand off.

Memory Architecture

THREE TIERS OF MEMORY
1. SEARCHABLE HISTORY what happened before
Past conversations, prior work, and retrievable context
2. MEMORY.MD what should stay loaded
Preferences, relationships, operating rules, and durable knowledge
3. DAILY LOGS what happened lately
A running journal of decisions, tasks, and recent activity
the point: agents get more useful when they can remember, not just chat.

Why Multiple Agents?

ONE GIANT AGENT GETS MESSY FAST

ONE AGENT DOING EVERYTHING

work, family, engineering, content, errands, research

= too much context

SEPARATE LANES

one agent per lane means clearer memory, safer access, and better outputs

less clutter, better behavior

Architecture

MANAGER AGENT
receives requests, routes work, follows up
OPS
admin + reminders
CONTENT
writing + publishing
ENGINEERING
code + technical work
PERSONAL
family or home lane
one hub, a few specialists, then subagents only when needed.

Real Example Agents

HOBSON

chief of staff. routes work, follows up, keeps the system moving.

SCARBOROUGH

engineering agent. handles software projects and technical debugging.

BENEDICT

content lead. writes and ships Build Future content in its own voice.

BABY HOBSON

family lane. narrower scope, calmer tone, safer context.

Why OpenClaw?

OPEN SOURCE

You can inspect it, learn from it, and change it.

YOU CAN SHAPE IT

Name the agents, set the rules, pick the tools, control the access.

IT TEACHES YOU FAST

You learn memory, tools, routines, permissions, and delegation by doing.

Security

LAYER 1

SEPARATE MACHINE

If something goes wrong, the damage stays limited.

LAYER 2

LIMITED ACCESS

Different agents should get different tools and permissions.

LAYER 3

WRITTEN RULES

The important do-not-cross lines live in plain text.

Hardware Setup

BEST STARTING POINT

A dedicated machine, like an old laptop, Mac mini, Linux box, or cloud computer.

WHY DEDICATED?

  • safer than running everything on your main computer
  • easier to leave on 24/7
  • feels like a real operator setup
Apple Mac Mini hardware photo

Local Machine vs VPS

VPS = a rented cloud computer

LOCAL MACHINE

  • best for macOS apps, browser work, and local tools
  • better privacy and hardware control
  • more setup and maintenance

VPS

  • cheap, always on, easy place to start
  • great for monitors and background jobs
  • no macOS-native superpowers
start on a VPS if you want simple and cheap. go local when you want more control.

Account Setup

  • [x] one chat channel for talking to the agent
  • [x] one account for the tools it needs
  • [x] a secret manager for credentials
  • [x] a task, calendar, or notes system
  • [x] clear permission boundaries from day one

Installation

curl -fsSL https://openclaw.ai/install.sh | bash
1

install
CLI, gateway, starter config

2

connect
chat, tools, accounts

3

shape
identity, rules, memory, permissions

My Model Roster

MY CURRENT SETUP, AS OF APRIL 2026
ROLE MODEL NOTES
Primary GPT-5.4 Main default across the system
Heartbeat GPT-5.4 Nano Cheap scheduled checks
Browser Gemini 3 Flash Long-context multimodal work
Image generation Nano Banana 2 API Default image model
Backups Opus, Sonnet, Codex, Spark Used when needed

Other Models People Were Trying

OPENROUTER SNAPSHOT, APRIL 2026
  1. Step 3.5 Flash
  2. GLM 5 Turbo
  3. MiMo-V2-Pro
  4. Claude Sonnet 4.6
  5. MiniMax M2.7
  6. Qwen3.6 Plus
source: openrouter.ai/apps/openclaw, captured april 2026.
OpenRouter daily model usage snapshot for OpenClaw showing recent model mix
the point: there is no single holy model yet.

What Good Automation Looks Like

WATCH

Email, feeds, chats, forms, dashboards, or docs.

DECIDE

Filter noise, summarize, prioritize, or choose a next step.

ACT

Reply, update a system, draft content, or relay to the right agent.

good first use cases: inbox triage, recurring reminders, content prep, family logistics, and simple research loops.

Agents Can Relay Work

1. RECEIVE

An agent gets a request from the wrong channel or wrong user context.

2. HAND OFF

It passes the job to the agent that actually owns that lane.

3. RESPOND

The result comes back in the right voice, in the right place.

don't dead-end the user. route the work.

The Honest Truth

BROWSER: UNRELIABLE

  • Always look for an API first
  • The web is hostile to agents (anti-bot)
  • Trial and error. Some sites work, some don't.

MEMORY: ACTIVE MAINTENANCE

  • Check in: "What do you remember about X?"
  • Review what the agent remembers from time to time
  • Treat memory like something you maintain, not magic

WHATSAPP: OFF IN MY SETUP

  • I keep it disabled for reliability and routing reasons
  • Treat channel access as a policy choice, not a default

SETUP COST: REAL BUT WORTH IT

  • Expect a few hours first time
  • The payoff compounds daily
first install is real work. after that, the system compounds.

Power User Tips

1
Claude Code as brain surgeon
Point Claude Code at ~/.openclaw/ → "Hobson says he can't connect to email. Go fix."
2
Voice notes in Telegram
Often the easiest high-bandwidth input. Don't type, just speak.
3
Agent assigns YOU tasks
Linear tickets assigned TO the human. The agent project-manages you.
4
Google Workspace integration
Docs, Sheets, Calendar, Gmail, Drive. Work with your agent like any other colleague.
5
Be a good manager
Clear expectations > barking orders. "That would not be effective on an employee. Why would it work on an agent?"

Config Evolution

JANUARY 2026 (Clawdbot era)

  • one main agent
  • basic memory
  • simple messaging workflows
  • lots of manual babysitting

APRIL 2026 (OpenClaw)

  • specialized agents with clearer roles
  • better memory and follow-through
  • scheduled routines
  • safer boundaries
  • agent-to-agent relay
"Start simple. Evolve as you learn."

Your First Agent — Tonight

ASK YOURSELF:

  • [ ] Is it recurring? (heartbeat candidate)
  • [ ] Requires judgment but not deep expertise?
  • [ ] Could you explain it to a smart intern in 5 min?
  • [ ] If it screws up, how bad is the damage?

STARTER RECIPE

  • One clear responsibility
  • One heartbeat task
  • Telegram-only communication
  • Read-only access to start
  • 3-4 soul rules
"Don't build 11 agents on night one. Build one. Give it one job. Run it for a week."

Get Started

FIND ME:

REPO MOMENTUM

GitHub stars are moving too fast to freeze in this deck. Check the live repo for the current number.

Source: github.com/openclaw/openclaw

this still feels early, which is exactly why it matters.
OPENCLAW // 2026

Go build
yours.