I built an agency in a day. Teaching it to say "I don't know" took the week

On the morning of 11 October I handed Claude a library of marketing playbooks and one instruction: build the world's best marketing company, and call it 4EVR.

By mid-afternoon it had sixteen agents in seven teams, a CMO called Vera, a way to hire people and a way for the whole place to learn from its own mistakes. By mid-afternoon. I have sat in meetings about the agenda for a meeting that took longer than that.

Then I spent the next six and a half days on things that never show up in a demo. This post is about those six and a half days, because that's where the one thing I actually want to tell you is hiding. I'm going to make you wait for it. Sorry. Old comms habit.

Context, for anyone who arrived here from somewhere sensible: I'm not a developer. I've spent fifteen years in communications. Daily newsletters, leadership messages, the internal side of a rebrand, every internal event, every single year.

And most of that job is not the writing. Nobody tells you this. It's turning can we do something about leads? into an actual brief, finding out what's true, getting the right people to look at it, and checking it before it goes out. The words are the last ten percent. The other ninety is chasing.

A friend had a pixel-art office full of little characters he could hand tasks to and watch work. In that set-up, he was God. I wanted to be God as well, obviously. Who doesn't. But mostly I wanted to know if a team of agents could run the whole loop and not just the bit at the end where words fall out.

First proper test. One line, pointed at a real company's website: get them more qualified leads.

The agents couldn't open the site. They treated a one-line request as a complete brief. And then they routed "leads" straight to cold email.

Reader, I have worked with agencies. This was the most realistic thing the software has ever done. Didn't read the site. Didn't ask a single question. Went straight for the cold email campaign. If it had sent me an invoice with a "strategy fee" on it I'd have assumed the build was finished and gone home.

So I did what you do with an agency. I wrote a process. A research step that reads the client's site and the wider web before anyone touches anything. A structured brief. And a rule I'm quietly smug about: anything the agents didn't read on a real page gets labelled Not verified. On the deliverable. Out loud.

Every claim either has a source or wears a little badge of shame. I would like this rule to be applied to LinkedIn, please, starting Monday.

The idea doing most of the heavy lifting is the boring one, and I love it. Agents look before they ask. The client's site first, then the internet, then any connected tools, and only after all of that, me.

Whatever they still can't find goes to an Inbox. Each agent gets three requests, each task gets eight, and a request that doesn't say what was already searched is refused. Refused! By the software! It just sends it back!

I want this on my actual email. Hi Tom, quick one, where's the logo? Refused. Please state where you looked. Please state that you opened at least glanced at the brand folder I've sent you four times.

And nothing waits on me. If I don't reply, the agent carries on with a clearly labelled assumption and moves on with its life. Which is, frankly, better manners than mine.

The team can hire. Only specialists built from real playbooks though (196 of them, from two open-source libraries), and never past the caps: 24 people in total, 12 hires, 3 per task. No empire building. There's even a little People & Talent team that writes the job descriptions and the performance reviews.

And the rating is not a vibe. Exceeds means 90% of your work got approved first time. Meets is 70%. Nobody gets rated at all until three pieces of their work have been reviewed. Letting anyone go needs my sign-off.

I have had appraisals with less maths in them. I have given appraisals with less maths in them. (Sorry. Growth is a journey.)

All of it lives in code. The agents can propose whatever they like, they just can't change the rules. Which is a sentence I'd happily print and stick on a few org charts I've known.

Ok. The thing I made you wait for.

The agents, the teams, the hiring and the learning took a morning and a bit. The rest of the week went on this list:

• stopping a task cleanly when you tell it to stop
• retrying without doing the work twice (an audit found a double-click on Retry ran a whole task twice, lovely)
• a $5 budget cap, after which it drops back to simulation instead of quietly spending your money
• error messages that tell you the truth

That last one hurt. Everything had been tested against fake versions of the model's API. 154 tests, all green, everyone very pleased with themselves, mainly me. Then the first live task sat in planning for a while and failed every single workstream.

The tool loop had been sending the model's own signed thinking back to it, which the real API rejects. And the planning step, instead of saying so, had been quietly falling back.

Quietly falling back instead of saying so. That's the most human bug I've ever shipped. I know people who have built entire careers on it.

The fix was small. The lesson wasn't. The clever part of an AI system is the part the demo shows you. The part you'd actually trust is the dull machinery nobody screenshots. Stop. Retry. Budget. Say what went wrong. That is the product. The office with the walking pixel people and the dog is the bit everyone notices first, and it's the bit that matters least.

(I'm keeping the dog.)

I don't know yet if 4EVR is any good at marketing.

Those 154 tests prove the plumbing works. They don't prove the plan it writes is a good plan. The live paths are tested against fakes, not real services. The system does measure whether its own lessons lift first-time approval, but on a sample far too small to brag about, so I won't. Everything it makes is a draft for a human to review, and the agents can't log in to your ad accounts, your analytics or your CRM. On purpose.

Bit of an anticlimax? Yes. But I'd rather tell you that than slap "AI-powered agency, 10x your pipeline" across a landing page. Plenty of people are already doing that, and none of them have a dog.

What is 4EVR HQ? An open-source team of sixteen AI agents in seven teams that research, plan, review, hire and learn, and you watch them do it in a pixel-art office. It's at version 0.9.0-beta.1, MIT licence.

Can I try it? Yes. It runs on your own computer and starts in simulated mode, which needs no API key and stamps everything SIMULATED so nobody mistakes it for the real thing.

Did you write the code? I directed Claude Code and argued with it about the details. Built 11 to 18 October 2026.

Does it work? End to end, yes. On real client work, not proven yet. See above. I'm not hiding it.

Why is there a dog? Why is there not a dog in your office.

Tom