Installing your own AI agent: what actually happens
What you end up with
Four pieces, all running on a server you control. They are separate on purpose: if one part breaks, the rest keeps working, and you can move or bill any of them without touching the others.
Your agent
The part that thinks. It reads, writes, uses tools, and works through tasks
on its own schedule. It gets a name and a written scope saying what it may touch
and what it must leave alone.
Where you watch it work
Every job the agent does lands here as a ticket, with the time it ran and what
it produced. You open one page and see the week.
What it remembers
A store of facts the agent writes as it learns and searches later, so it stops
forgetting between runs. It holds three entries on the day we leave and eleven a
week later. This is the piece that makes the agent better in month six than it
was on the day we handed it over.
Your customer database
Contacts, companies and deals, held on your server, that the agent reads and
updates. Most businesses pay a monthly subscription for this part on its own.
Why four parts. One app that does a bit of each is cheaper on
day one and charges you monthly for the rest of its life. The four parts cost more
to set up once, then keep working with no subscription, and you can replace any one
of them without touching the others.
What you bring to the table
Most of the install is our work. Five things need you, and only the first is
technical.
| What | Why it needs you |
|---|---|
| A server | Any Proxmox host, or a virtual server you rent. Four small containers, 26 GB of memory between them. If you already run a server, this costs nothing extra. |
| A sign-in for the agent's model | The agent calls a model over the internet instead of running one on your hardware. That call is the one ongoing cost, and it is metered by use. |
| An email address for it | So it can send you the morning note. We recommend it drafts and never sends without your OK until you say otherwise. |
| Your scope decisions | What the agent may touch, what it must never touch, its working hours and its timezone. Twenty minutes on a call covers it. |
| Access to your records | If you already use a CRM, we decide whether to bring it across. If not, we stand up the one above and load what you have. |
If you have none of this, the server is the only purchase, and it is a monthly
bill you can cancel.
The week, day by day
Fit
We look at what you already run and say whether this is worth doing. If it is
not, you hear it on that call. We agree the scope, the name, the hours and the
timezone.
The worker goes in
The agent is installed on its own part of the server, given its key, its name
and its written scope, and taught your setup. It gets its own map of your
network, built from a real look at what is there.
Memory and the board
The memory store goes up and the agent is wired to it. The board goes up next,
and the agent is connected so its work becomes visible. It is the fiddliest day,
because the parts have to agree with each other.
Your records
The customer database is built and, if you have existing records, loaded. The
agent gets the key it needs to read and update them.
It starts working on its own
Its routines go in: a morning note, an end-of-day write-up, a nightly backup,
a weekly review. All in your timezone, all visible on the board. Then we stop
helping and let it run.
We watch it do the job
We wait for a routine to fire with nobody touching it, then read what it
produced. If nothing fires, or what it produced is wrong, we fix it before
handover.
Handover
You get the logins, the server map, the schedule, and a written list of
anything still outstanding. Our own access is removed. You keep the
passwords.
How we prove it works before we leave
Each claim below is checked on the machine and written down with its result, bad
ones included.
| Claim | How it is shown |
|---|---|
| The agent is running | It reports its own version, and the service survives a restart. |
| You can log in | A correct password gets in, a wrong one does not. |
| It remembers | We count what is in the memory store, have the agent write facts, and count again. The number moves or the claim is false. |
| You can see its work | Its comment appears on the board under its own name. |
| It works unattended | A routine fires with nobody in the loop, and leaves a record plus the thing it produced. |
| Your data survives | A dated archive is written every night. Fourteen days of them are kept, the oldest rolls off, and nothing you put on the box is deleted by it. |
| It comes back after a reboot | Every service is set to start itself, and we confirm it does. |
A real install
The most recent build went onto a customer's own three machine cluster, on
hardware he already owned. Nothing of his was replaced and nothing was exposed to
the internet. Four services went in, the agent ran her routines unattended, and our
access came out before we called it done.
The whole job is written up, faults and open items included: read the case study.
What was still open when we handed over
- His agent's email address was assigned but not connected, and the intended
rule was draft-only. - She held no key to the customer database, so she could not file a contact.
She reported that herself, in her own end-of-day note. - Two stale entries sat on the board. Clearing them is his call.
What you own at the end
- The server. Your machine, your network, your bill.
- The data. Every fact, contact and record sits on your
hardware. Nothing is mirrored to us after handover. - The logins. You are the administrator. We hold a scoped key
while we work, and it stops working when we are done. - The scope. The rules saying what it may and may not touch are
written on the box and you can change them. - The ability to take it elsewhere. It is ordinary software on
an ordinary machine, with none of it locked to us.
What this does not include
The install does not cover editing your website, running your advertising, or
doing business research and strategy for you. We build and hand over the system.
What it works on afterwards is your call.
One cost to keep in mind. The agent calls a model by the use.
Left to run all day on long tasks, that bill grows. We set its routines to cheap
work by default and show you how to watch the spend.