A hospital bed kept its serial number for life. Parts came and went. When a motor burned out, the manufacturer sent a replacement, we swapped it in, and the dead one went back so they could confirm what had actually failed. The bed's number never changed. The new part's number got attached to it.

So the record was never one number. It was three things. Which unit this is. What is on it right now. What was swapped, and when.

I keep coming back to that when people talk about logging AI.

Most of what I see logged is the conversation. What was asked, what came back. That tells you the output. If something unexpected happens, it does not tell you the cause, because six weeks later nobody can say what was running at the time.

An agent needs both halves of the number. Which agent, and at what version.

The agent is the part that persists. The one that reads your invoices is the same agent in March and in September. Give it a name and keep the name.

The version is what was actually in it on a given day. Which model and which release of it. Which instructions. Which tools and permissions. What data it could reach.

Then the one people skip. What changed, when, and who approved it.

That one already has a name, and most work has some version of it. A change log on a tax return. A revision history on a drawing. A change record before anything goes into production. The rule is the same everywhere. You do not change the thing without writing down what you changed and who signed off.

An agent is that thing. Its instructions, its permissions, the data it can reach are all changes to something already in use, and they usually happen in a chat window with nothing written down.

Here is the part that should bother you. A bed motor does not replace itself. Somebody has to go out there with a wrench, and that leaves a record whether anyone means it to or not. The model underneath your agent can be changed by the vendor, while it is in use, with no change record anywhere on your side. Your configuration says exactly what it said yesterday. The behavior does not.

When that motor failed there was a part to send back. With an agent there is nothing to return but the log, so if the log does not say what was running, you have sent back an empty box.

Two things to do with that.

The floor, and you control all of it. Write down your half. Which tool. Which account. Who set it up. What you changed and when you changed it. If you already have a one-page AI rule, this is one more line on the same page.

The ceiling, and you have to ask for it. What version am I running, and how will you tell me when it changes. That is a fair question to put to any software vendor, and how they answer tells you something either way.

Six weeks from now something an agent did is going to need explaining. The question is whether you will be able to say what it was running at the time.