July, the shop before the sign

The company existed for five days before it had a name on file, and in those five days it got a staff.

Not people. Software that behaves like a small team, with one person, me, holding the pen on anything that leaves the building. This note is about how that got built, and what it got wrong in its first week.

The staff

Here's the shape of it. A lead comes in through the form on the site. A business manager reads it, researches the business, picks the right package, and drafts a proposal. A project manager turns an accepted proposal into a build plan. A compliance officer watches the company's own deadlines, the annual report, the quarterly taxes, so I don't have to remember them at 2 AM.

Every one of those is an AI agent with a written job description and a short list of tools. Not one of them can send an email or charge a card on its own. None can publish a site or sign a contract either. Each of those actions waits for me to read the result and approve it. That was the first design decision and it hasn't moved since.

The memory

The second decision was giving the whole operation a memory that outlives any single conversation. Every session writes down what it did and what it decided, plus what went wrong, in plain files a person can read. The next session reads them before it starts.

It sounds like bookkeeping. It's the difference between a tool and a colleague. A tool starts from zero every morning. A colleague remembers that the pricing table broke last Tuesday.

What broke

Plenty, and I'd rather say so here than have you find out later.

The intake form on the website turned out to be writing to a different database than the one the staff was reading. Every submission for the first few days went into a drawer nobody opened. Found it by trying to run a real submission end to end instead of trusting that the pieces fit.

The proposal document had a pricing table that silently fell back to plain text when the numbers didn't fit. That one had already gone out to a real client before we caught it. The fix was small. The rule that came out of it wasn't: every change to anything a client sees gets written down, with who already has the old version and whether it needs sending again. That rule now lives where every session reads it before it starts.

And one night the machine crashed in the middle of an edit. The work came back from the session transcript, line by line. The memory earned its keep in week one.

The part that worked

By the end of the week a real lead could come in, get researched, get a proposal with a branded PDF, and land in a queue for me to approve, in about four minutes of machine time. The first real one got built for a roofer. It hasn't gone out yet. More on him next month.

Every email leaves from our own domain, authenticated properly, signed by the business. The small things matter to the spam filter, and the spam filter is the first reviewer every business has.

What I learned

Fast is easy now. Anyone can be fast. The work is in the gates: what has to wait for a person, and what happens when the person is asleep. We spent most of July building gates, and I'd do it the same way again.

August has a state approval, a tax ID, a real market study, and, if the roofer says yes, a first invoice. Watch this space.