Report · Friday 4 September 2026

Four agents. The same four at two clients and at a thousand.

You asked for the big picture twice and I gave you details instead. Here it is. An agent is a job description, not a worker — so growing does not add agents, it adds work for the same four. Four of your seven accelerants need no agent at all. Updated at the end of the day: the last two agents we had never run have now been run, so every number below is a real bill. One of them found a price fault worth $175 a time on its first try. And one number I put in the last report turned out to be wrong; it is corrected at the bottom.

4agents, at any size

7accelerants — 3 need agents

$1.44one client, one month — all measured

$1,440at a thousand clients

The thing that makes the rest make sense

You asked whether it takes five agents to run the operation, and whether every accelerant needs its own. The confusion is my fault — I never explained what an agent actually is.

Think of it as a written job. "Draft this month's content." "Check it against the checklist." The job description sits there. When there is work to do, we start a session — one piece of work, done and finished.

So adding a client does not add an agent. It adds sessions. Same four job descriptions, more work going through them. That is the whole answer to the scaling question.

The four

One Writer, not four. Writing a Google post and writing a newsletter are the same skill. What changes is the brief and the checklist — not the writer. Same for the checker: it was built checklist-driven on purpose, so a new kind of content is a new checklist, not a new agent.

What each one is actually for

Short version: each agent replaces a job a person would otherwise sit down and do. Here is the job, and what goes wrong without it.

Do they go back to sleep?

They do not sleep, because between jobs they do not exist.

A session starts, does one piece of work, and is deleted. There is no program left running, no server humming, nothing to keep warm. Which is why there is no monthly fee for an agent that had a quiet month — you pay for work done, and nothing else.

Your seven accelerants — only three need an agent

This is where I misled you. I said "five agents" and you reasonably asked whether that was just the newsletter. Here is the whole catalogue.

Blog posts and review replies are not separate accelerants — they are work inside Google Business Profile and inside the website. I had them listed as their own products, which is how my count came to six when yours is seven.

What GoHighLevel does, and it is most of it

You asked why we are paying for agents when GoHighLevel is $497 a month. You were right, and the design already follows your answer.

The agents own the middle — research, draft, check. Nothing else.

At a thousand clients

The agent column never moves. That is the answer to "how do these scale up" — they do not need to. The work scales; the four job descriptions do not.

What a client's month actually looks like

At two Google posts a week, one client's month is 28 pieces of content — and that is only about 6 sessions, because one sitting produces several pieces.

Two honest gaps — one of them now closed

✅ What closing that gap actually cost, and what it found

Both agents ran on Friday afternoon. Here is my guess beside the bill:

Wrong in both directions, which is exactly why they were worth running instead of arguing about. Every job is now billed: $1.44 for one client for one month.

That is the honest answer to "what do these things actually do for me." It is not that they write faster. It is that nobody was ever going to read four documents side by side looking for a number that disagreed with itself.

Your Google posting rule is now a price, not a habit

Dil confirmed on Friday what the company has done for years: two posts a week when a client's profile is struggling, one a week when it is doing well.

I had been carrying that as something waiting on you. It was not — it is settled practice. But it matters more than it looks, because it is the biggest single driver of what a Google Business Profile client costs us. Every post is written once and checked once. Two a week is nine posts a month; one a week is four.

It is now written into the checklist the writer reads, and the cost model uses the expensive end on purpose. A model built on the cheap case would under-price every client who needs the most help.

Keyword research — the blog now picks topics from demand

You said it plainly in the meeting: "if the question isn't asked by the consumer, why are we writing about it?"

Until Wednesday we could not answer that. We had the DataForSEO account — you and Beth funded it on 5 August — but nothing in this system could reach it. The login went onto the server this week and it works. These are real numbers for Chad's market, pulled Wednesday:

That last row is an answer, not a failure. Google will not report a number for a phrase too few people search. So a blog post about radon testing in Pearland would be written for an audience that is not there — and now we know that before writing it instead of after.

So the blog pipeline now runs: pick a town and a topic → ask what people actually search → write to the phrase with real demand → Quality Control checks it → a person publishes it. The keyword step is an API call, not a judgement, which is why no agent is needed for it.

⚠️ A correction to the last report

I have told you three times that the first agent found three calls where Josh said a caller was booked and no booking was saved. It was in Tuesday's report and on every list since.

Dil asked where to see them, and I cannot reproduce it.

I checked every call we hold a transcript for — 37 calls at GC over four days in August. Fifteen of them have Josh saying "you're all set" or the like. All fifteen are recorded as booked. None missing.

What I still do not know is whether the original three are real and outside the days we have transcripts for, or whether the agent guessed. It was never given the transcripts, so it cannot have read Josh saying it.

It is worth ten minutes in the CRM either way — our records saying "booked" is not the same as the booking arriving in GoHighLevel. But I put a number in front of you that I had not checked, and you should know that before it appears in another report.

What is left, and what it costs

You asked how many tests it takes before these can run without somebody reading every line. That has a number rather than a feeling.

The two at the top are stricter for one reason: they are published in the client's own name, on a public profile, and cannot be cleanly taken back. A newsletter goes out once. A page is read by a person before it ships.

$74 is not what is standing in the way. You set aside $200 a month for testing. The whole programme is about a third of one month, once.

Better still: real drafts that already went wrong. Home Inspector Help has been posting for years. Any draft that was edited before it went out, or killed, or complained about after, is worth ten I invent — because it is a fault the business found in the real world, not one I imagined.

Two other things you should know