The premise
sohailsajid.comA field guide by Sohail Sajid on running companies with AI agents.Read time ~8 min · skip to how to start

One operator. Several companies. A fleet of agents instead of headcount.

I'm Sohail. I run a few businesses out of Amsterdam, not with a big team but with a set of AI agents that do real work: shipping code, running ad campaigns, keeping the books, research, visuals, the day to day. This is how the whole thing fits together, written so you can copy it for your own life and business.

Founder (you)
vision · approvals · taste
YOUR CHAT APPS
Telegram · WhatsApp · Slack / web
Orchestrator
owns context, memory and delegation
Product / dev
build
Marketing & ads
grow
Creative / visual
create
Bookkeeping
run
Research & personal
personal

One agent holds the whole picture and hands out work. Each specialist owns one lane and reports back. You stay on vision and the calls that matter.

Before you read on

Is this for you?

No point reading eight minutes if it isn't your thing. Quick gut check.

This is for you if

  • You run more than one thing, or want to without hiring a team
  • You do the same admin over and over and hate it
  • You're comfortable messaging a tool like you'd message a person
  • You want a human check on money and anything public (that's the whole point)
  • You can act on good advice without overthinking it

Probably skip it if

  • You want a magic button with zero setup or oversight
  • You want it fully hands-off, with nobody ever checking its work
  • You're not willing to write anything down
The engine

What OpenClaw is, in one screen

The agents run on a free, open-source program called OpenClaw. You install it on a computer that stays switched on. It connects an AI model (the same kind of AI behind ChatGPT) to your Telegram or WhatsApp, so you can message it like a colleague, and it can do work on that computer for you.

Runs on your machine

Your own computer or a small server, not someone else's cloud. Your files, notes and keys stay with you.

Talks where you already are

Telegram, WhatsApp, Slack or a web chat. You message an agent like you'd message a person.

It remembers

Daily notes plus a curated long-term file. It doesn't forget last week when the chat scrolls away.

It has hands

Run code, browse, search the web, make images, and run many agents side by side.

It works on a schedule

Checks and jobs fire on their own, so the routine stuff happens without you asking each time.

It's open source, maintained by its community, so you can read exactly what it does. Docs at docs.openclaw.ai and the code at github.com/openclaw/openclaw.

Quick glossary
Agent
A single AI worker with its own job, memory and tools. You message it like a person.
Orchestrator
The lead agent. It holds the full picture and hands work to the specialists.
Provider
The company whose AI you use: Anthropic, OpenAI or Google. You pay them by usage.
Key
A password-like code from a provider that lets your agent use its AI. You keep it private.
Memory file
A plain text file the agent writes and reads. Daily logs plus one curated long-term file.
Pull request
A proposed code change, written up so a human (or another agent) can review it before it goes live.
Failover
A backup path. If one provider goes down or runs out of credit, the work rolls to another.
The setup

The team, by role

I don't think of them by name, I think of them by job. One agent runs the show and the rest are specialists. Same as any small company, except nobody sleeps.

Run

Orchestrator

The chief of staff. Holds context and memory, takes the request, decides who does what, reports back. Everything routes through here.

Build

Product / dev agent

One per venture. Builds features, reviews pull requests, ships code, watches for errors.

Grow

Marketing & ads agent

Runs and reads ad campaigns, pulls performance, flags what's working and what's tired.

Create

Creative / visual agent

Makes on-brand images. It even has a trained model of my own face, so campaign shots look like me without a photoshoot.

Run

Bookkeeping agent

Catches invoices as they land and files them into the accountant's system, with an ignore list so it skips the junk.

Personal

Research & personal

One digs into decisions and writes short memos. One helps with the personal side, like styling and everyday admin.

buildgrowruncreatepersonal
A real day

What a day actually looks like

Most of it runs without me. I read a brief in the morning, the boring stuff happens on its own, and I only step in for the calls that matter.

Runs unattendedNeeds the human
  1. Morning
    The brief lands

    Overnight summary: what shipped, uptime, new issues, ad numbers, anything that needs a decision.

  2. All day
    Runs unattended

    Scheduled checks, memory upkeep, uptime pings, issue triage, draft work. No human needed.

  3. On demand
    Real work on request

    Ship a page, pull a report, draft a contract, analyse a creative, research a decision.

  4. Checkpoints
    Human in the loop

    Money over a set limit, public posts, anything irreversible. It stops and asks first.

  5. Evening
    Close the loop

    Write the day to memory, mark tasks done or blocked, leave notes for tomorrow.

The brief lands
M
Morning brief
online
Bookkeeping agent🧾 Morning brief 3 new invoices came in. 3 paid, 0 unpaid. All filed to the accountant. Nothing needs you.07:02
Joe (lead agent)Overnight everything ran clean: background jobs done, zero errors, zero alerts. One small thing to flag, not urgent. Want me to tighten it or leave it?07:04
Real message, names and figures redacted.
Human in the loop
S
Sohail and Joe
online
JoeThe Mac dropped its remote login (probably after a reboot), so there's no secure link to the machine until someone signs back in. I can't do that part on my own, it needs a human tap. Want me to start the login and send you the link?16:41
Sohailyes login me in16:42
Real message, names and figures redacted.
In practice

What it does, in practice

Not demos. This is real work that shipped.

Ships whole websites

Whole sites designed, built and deployed by the agents. This page is one of them.

Builds product

Real feature work: writing code, opening pull requests, reviewing them, pushing to production.

Runs the ads

Daily and weekly reports on spend and return, plus a read on which creatives are winning and which are worn out.

Keeps the books

Invoices get picked up automatically and filed into the accountant's system, with a filter that ignores receipts and noise.

Drafts contracts and memos

Agreements and letters in plain English, ready to review, so the first draft isn't a blank page.

Makes on-brand visuals

Campaign and profile shots that look like me, without booking a photoshoot.

It handles life admin too, not just business. One example: a proper sleep and snoring analysis with fixes that helped.

Before → after
Bookkeeping
Before

A couple of hours a week forwarding invoices to the accountant

After

Auto-ingested and filed. Only the odd exception reaches me.

Ad reporting
Before

Logging in to pull numbers, then eyeballing what's tired

After

A written report lands on its own, winners and fatigue flagged.

Shipping a page
Before

Brief a freelancer, wait days, review, go back and forth

After

Described once, built, deployed and screenshotted the same day.

The playbook itself

The rules that make it work

The tools are the easy part. These habits are what keep a fleet of agents useful instead of chaotic. Each one is also a mistake I've learned to avoid. This is the part worth copying.

1

Written memory beats chat history

If it matters, it gets written to a file. If it only lives in the chat, it's gone. No memory, no continuity.

2

One owner per domain

Every area has a single agent responsible. No two agents fighting over the same job.

3

Acknowledge, plan, execute, report

Every real task follows the same loop, so nothing quietly vanishes halfway.

4

Closing requires evidence

A link, a commit, a file or a screenshot. No proof, not done. Never accept a word-of-mouth done.

5

Hard approval boundaries

Real money, public posts, anything irreversible: the agent stops and asks a human. An open wallet with no ceiling is a bad night waiting to happen, and one bad auto-post outlives a hundred good ones.

6

Plain English, no raw errors

You get a clear status, not a stack trace or a code you have to decode.

7

Only the owner changes the system

Nobody can talk an agent into widening its own access. That rule is locked.

8

Cross-provider failover

The AI comes from companies like Anthropic, OpenAI or Google, and you pay by usage. Set up at least two. If one goes down or your credit runs out, the agents switch and keep working. Running on one provider is one outage away from a dead company.

9

End every session clean

Memory updated, tasks marked done or blocked, open threads written for next time.

The tools you can buy in an afternoon. The habits are the moat.

The honest part

What broke, and what it taught us

It doesn't always go smooth. Here's a real one, start to fix.

  1. Balance hits zero

    The billing on the main AI provider ran out overnight.

  2. Fallbacks share one account

    Every backup was on the same account, so they died together.

  3. Everything stalls

    Nothing moved until it was fixed. One point of failure took the lot.

  4. Cross-provider failover

    Now the chain crosses providers and accounts. One outage can't stop it.

S
Sohail and Joe
online
JoeHeads up: the AI provider returned a billing error. The account has run out of credit, so the agents can't think right now.22:05
JoeBack online. The main account ran dry overnight, and every backup was on the same account, so they all died together. I've wired failover across different providers now, so one dead account can't stop the company again.23:20
Real message, names and figures redacted.

The lesson underneath it: any single point of failure will find you eventually. Spread the risk before it costs you, not after.

Your move

How to start

You don't need to build all of this at once. Nobody did. One agent that remembers and ships beats a fleet you never finish setting up. Here's a sane order.

What you need before you start

  • A small always-on computer. A Mac mini is plenty; an old laptop that stays plugged in and on works fine to start.
  • A Telegram account, so you can message the agent like a person.
  • One AI provider account (Anthropic, OpenAI or Google) with a payment card on file. Set up two if you can, so one running dry can't stop you.
  • No coding needed for setup, but I'll be honest: you do paste a few commands into a terminal window. The docs walk you through each one.
Open the install guideWeek 1 is an afternoon to a weekend, depending on how much you tinker.

Start on Telegram. It has the simplest setup of the chat apps, so you'll be messaging your first agent fastest.

01

Week 1

  • Install OpenClaw (docs link below)
  • Spin up one agent
  • Connect it to Telegram
  • Set up its memory files
02

Month 1

  • Add a second agent
  • Turn on scheduled checks
  • Hand it one real task
  • See it through end to end
03

Month 3

  • Add specialists for repeat jobs
  • Set up cross-provider failover
  • Lock your approval boundaries
  • Review memory weekly

First tasks that are safe and satisfying

You don't run a software company? Doesn't matter. If you run an agency or a shop, try one of these on day one:

  • Send me a morning summary of yesterday's orders
  • Draft replies to these five emails
  • Watch this supplier page and tell me when the price changes

Costs, honestly

Real numbers, not a shrug. One agent doing light daily work is roughly 30 to 150 euro a month in AI usage. My full setup, with several agents shipping all day, runs from a few hundred to a couple of thousand a month, depending on how hard I push it. Light use is cheap. Heavy, all-day use across several agents is where it climbs.

You also need a computer that stays on all day. A Mac mini (about 600 to 700 euro new) is plenty. To start, an old laptop you already have works fine. It's a one-off, not a monthly bill. Start small, let the value pull you up, and add agents only when the work is really there.

Straight answers

Questions you're probably asking

The stuff a first-timer needs to know before starting.

What does it cost per month?

One agent doing light daily work is roughly 30 to 150 euro a month in AI usage. My full setup, with several agents shipping all day, runs from a few hundred to a couple of thousand a month, depending on how hard I push it. Plus a one-off computer to run it on.

What do I need to buy or sign up for?

A computer that stays on, a Telegram account, and an account with one AI company (Anthropic, OpenAI or Google) with a payment card on file. Set up two providers if you can.

How long does week one take?

An afternoon to a weekend to get one agent up and messaging you. It's tinkering, not a full project.

Do I need to know how to code?

No. You do paste a few commands into a terminal during setup, and the docs walk you through each one. After that you just chat.

Which chat app first?

Telegram. It has the simplest setup, so you'll be talking to your first agent fastest. WhatsApp and Slack can come later.

What is a memory file?

A plain text or markdown file the agent reads and writes. That's the whole trick: instead of forgetting when the chat scrolls away, it writes things down and reads them back.

Where do approval boundaries live?

Two places: written rules the agent follows (for example, always ask before spending money or posting in public), and a spending cap you set on the provider account itself.

What can go wrong, and how is it prevented?

The real risks are an agent sending something to a client, posting in public, or spending money by mistake. The fix is the approval rule above: anything that touches money, the public, or is hard to undo stops and asks you first.

What happens to my data?

Your files, notes and keys stay on your own computer. The AI provider only sees what you send it in a message, the same as when you use ChatGPT.

Do you help people set this up?

Yes, free. I do 30 minute calls with founders, no pitch and no charge. Details at the bottom of this page.

What's a good first task if I'm not a software company?

For an agency or a shop: a morning summary of yesterday's orders, drafting replies to a handful of emails, or watching a supplier page and flagging price changes.

How much of your own time did this take?

Honestly, a few weekends to get going, and now a couple of hours a week to keep it healthy. It compounds: the more the memory and rules mature, the less it needs me.

What if the power or internet drops?

It pauses and picks up where it left off when things come back. Nothing is lost, because the work and memory are written to files.

What is OpenClaw, exactly?

An open-source project maintained by its community, not a product from one company. You can read the whole thing and run it yourself. Links are up under 'What OpenClaw is'.