Report · Monday 1 September 2026
In our last meeting you asked a simple question: could a few AI helpers from Anthropic replace the forty-odd agents we are paying for and not using? The answer is yes. This report explains what they are in plain words, which of our seven accelerants they actually help, and what they cost — including one real example we measured rather than guessed at.
Read this first Nothing has been built and nothing has been spent. This is a look-before-we-build report. No agent is running, no customer has been touched, and no money has left the account.
Everything here needs your yes before it happens.
Yes, we can build our own AI team with Claude. Anthropic has already built this. It is in our account today. The page is empty only because we have not made anything yet.
We do not need a new server. Anthropic runs the computer. This is the opposite of how Josh works, and it is good news for us.
Websites should stay exactly as they are. We measured this. Building one our usual way costs nothing extra. Letting an agent do it costs money. So we keep doing what we do.
The agents earn their keep on monthly work. Jobs that come back every single month, where nobody wants to sit and watch. The newsletter is the best fit we have.
Every one of your worries from the meeting has an answer. Spending limits, going stale, agents not talking to each other, going off and spending our money. All five are covered below.
One paragraph, then an example. This is the whole idea.
Right now, when we use Claude, a person has to sit there and ask it for things. It waits for us. If nobody asks, nothing happens.
An agent is different. You set it up one time. You give it a job, a schedule, and a spending limit. Then it wakes up on its own, does the work, and stops.
Think of it this way Today we have a helper we have to call every time we want something.
An agent is a helper who shows up every Monday, because that is the deal we made. Nobody has to remember. Nobody has to call.
No. And this is the part that surprised me most.
Anthropic runs the computer. When a job starts, they build a fresh work space for it. When the job ends, they take it down. We do not set anything up. We do not pay for a machine sitting idle waiting for work.
We can also let an agent look at our code folder when a job needs it. That is a choice we make job by job, not a rule. Most jobs will not need it.
One warning, and it matters This does not mean we can drop Rose Hosting.
Josh still needs that server. So does the phone line, the booking system, and both dashboards. This is extra. It replaces nothing we run today.
I am saying this plainly because we cancelled a server last week, and I do not want anyone reading this report and thinking we can cancel another one.
You asked for a real number. So I measured our own site instead of guessing.
Our Booked Solid website is 31 pages and about 1.3 million characters of code. That is a full, real website. Not a small one. It is a fair thing to measure because we built it ourselves and we know exactly what went into it.
There are two ways to build a site like that.
| How we build it | What it costs us | What we get |
|---|---|---|
| A person and Claude Code (how we do it today) | $0 extra | The finished website |
| An agent builds it alone | About $40 to $90 | The same finished website |
The first one is free because it is already covered by the monthly plan we pay for. The second one is billed job by job.
So the rule is simple Keep building websites the way we do now. A person and Claude Code, on the flat monthly fee.
We should not pay extra for a job we can already do for nothing.
That is not a knock on agents. It is just the wrong job for them. A website gets built once. An agent earns its keep on work that comes back every single month.
The other side of the same coin.
Take the newsletter for one client, every month. Find the content. Write the homeowner letter. Write the agent letter. Write the report at the end.
| Job | Rough cost |
|---|---|
| One client, one month | About $5 |
| Fifty clients, one month | About $250 |
That is my estimate, not a bill Nobody has run one of these yet. I worked the number out from what Anthropic charges, not from a job we have actually done.
But we do not have to guess for long. We set a $5 limit, run it one time for Chad, and read the real number off the screen. That test costs us five dollars at the very most, because the limit stops it.
I would rather measure it than argue about it.
There is no monthly fee. Three things get charged, and nothing else.
| What | Rate |
|---|---|
| The work itself | Charged by how much the AI reads and writes |
| Web searches | Ten dollars for a thousand searches |
| Work space time | Eight cents an hour |
And one saving worth knowing about. When a job runs the same way each month, the repeated part comes with up to 90% off. The newsletter job does the same thing fifty times, so this matters to us more than most.
These are your own words. Each one has a real answer, not a promise.
That was true of the old setup, and it was a fair complaint.
Here an agent is a saved setting, not a personality that drifts. You build it once and it stays put. When you want to change it, you change it — and any job already running keeps using the old version until it finishes.
Nothing rots while nobody is looking.
These share one notebook. What one agent learns, the others can read.
So the agent that writes the newsletter can see what the agent that checks the work already found, without anybody carrying a message between them.
This is the one I would want you to look at hardest, because it is the strongest answer of the five.
Every job takes a hard dollar limit. You set it in advance. When the job reaches that number it stops and waits for you.
It does not die. It pauses. Everything it has done so far is kept. You decide whether to give it more money or leave it there.
That is a real stop, built into the system by Anthropic. It is not us remembering to be careful.
We list the websites an agent is allowed to reach. Everything else is blocked before it starts, not talked out of it afterwards.
We can also cap how many searches it is allowed to make in one job. Our content puller already works this way — it is capped at eight.
We can make an agent stop and ask before it does anything that spends money or reaches a customer. It sits there and waits for a yes.
So on ads, for example, the agent can read the numbers, write the report and draft the changes — and then stop. A person presses the button that spends the money. Every time.
I went through all seven rather than picking the ones that sounded good.
| Accelerant | Does it help? | Why |
|---|---|---|
| Client & Agent Newsletter | Yes — the best one | Half built already. Finding the content, writing both letters and doing the monthly report are all agent work. |
| Paid Ad Management | Mostly | It can write the reports, draft the ads and spot wasted spend. A person still approves the money. |
| Google Business Profile | Would work | It fits well. But you closed that offer, so we wait until you decide to reopen it. |
| Local Microsites | For upkeep | Good for keeping town pages fresh every month. Build the sites themselves in Claude Code, free. |
| Trusted Home Pro Network | For the writing | Good for the directory and the vendor pages. The chat that sells to vendors is a different tool. |
| Converting Website | No | Build it in Claude Code. It is already free for us, as the cost example above shows. |
| Authority Builder | No | It is a one-time job, not something that comes back every month. |
Four of the seven are a real fit. Two are not, and I would rather say so than pad the list. The seventh is Google Business Profile, which fits fine but is waiting on your decision, not on the technology.
You can see the cost per client, per month. That is the number Beth needs to price things properly. Right now she is working it out by hand. This produces it on its own, as a by-product of doing the work.
Nothing reaches a customer, and nothing spends money, without your yes. The rule we already follow becomes something the system enforces, rather than something everyone has to remember.
One helper serves client number one and client number fifty. We grow without hiring.
Four good helpers instead of forty stale ones. Which was your point in the meeting, and it still stands.
One agent. The one you asked for yourself.
In the meeting you said we need "some kind of AI quality control that says it did it." That is the one I want to build first, and I want to build only that one.
Here is why that one.
| Reason | What it means |
|---|---|
| It only reads | It cannot send, post, or spend anything. There is no button on it that costs money. |
| Low risk | The worst thing it can do is write a wrong report. Nobody outside the company sees it. |
| It proves the setup | The schedule, the spending limit and the record all get tested on a job where being wrong is cheap. |
| You need it anyway | It is what lets us serve 500 clients without you and Beth checking every one by hand. |
In this order
Nothing starts without you