The harness is the new desk

I connected everything to Claude in an afternoon. Inbox, calendar, CRM, Stripe. Then I asked it the obvious question — who should I contact this week? — and it gave me a genuinely good answer. Names I'd forgotten. Reasons attached. Sorted into something that looked a lot like a plan.

Then I closed the laptop and contacted nobody.

I want to explain that gap, because it isn't a story about a bad model. It's the same assistant that drafts my proposals and writes my release notes, and it did exactly what I asked, well. The problem sits downstream of the answer, and once you see it you can't unsee it in any of your own tools.

Your agent writes your code, your emails and your documentation. It has no idea which customer to call.

Not for want of data. I'd just handed it the data.

Ask a developer why their AI feels so much more useful than yours and the answer is boring: their work already writes itself down. Every change, every version, every thing that broke and what fixed it — recorded automatically, because that's how software has been built for twenty years. So the agent turns up and finds the entire history sitting there, readable, without anyone having done anything for its benefit. It walked into a house somebody else had already furnished.

Open the same assistant to do sales and you get a text box.

Sales has no equivalent, and not for want of software. It's the opposite problem: the state is scattered across twelve products that don't know about each other. Who this person is, in the CRM. What they actually said, in the inbox. Whether they turned up, in the calendar. What they paid, in Stripe. Whether they still use it, in the product. And the part that decides everything — whether they're worth an hour this week — is in my head, and only on good days.

So when you point a capable agent at all of it, you get precisely what a very good new hire would produce on day one with every password and no manager: a large, correct, undifferentiated pile.

Three things, none of them clever

A harness isn't the model and it isn't the connections. Strip it back and it's three unglamorous properties, all of which engineering has and stopped noticing years ago.

Somewhere to put the result. A chat log is not a pipeline. If yesterday's conclusion doesn't exist this morning, you re-derive it this morning — and re-deriving is most of the cost. Worse, you re-derive it identically. The three opportunities you consciously looked at and decided to skip come back tomorrow looking exactly as urgent as they did today, because nothing recorded that you'd already made that call.

An opinion about order. Not a score — an order, with a reason attached, that survives you disagreeing with it. Twenty ranked opportunities is still twenty decisions, and I've spent fifteen years watching what happens when you hand a busy person twenty decisions: they make none of them and feel bad about it. One next action is a decision somebody already made. That difference is the whole distance between a tool you open and a tool you use.

The ability to act, not advise. The gap between "you should reach out to Marta" and a drafted message carrying the specific history — what she bought, what went wrong in month four, who else was on that call — is where the hour actually goes. Advice is free, and worth about that.

Take any one of the three away and the whole thing collapses back into a very articulate chatbot. You don't need to run that experiment to know it. You've been living in the result.

The two-minute test

Here's how to find out which one you've got, and it costs nothing.

Ask your agent who you should contact this week. Read the answer — it'll probably be good. Then close the session, open a fresh one tomorrow, and ask the identical question.

If the second answer doesn't know you already dismissed three of the first five, doesn't know you emailed one of them yesterday, and can't tell you what changed overnight — you don't have a harness. You have an extremely capable consultant who forgets you between meetings, and you are the only thing here that remembers.

That's worth sitting with, because it explains something that otherwise looks like a discipline problem. Every renewal that gets caught, gets caught because I remembered. Every dormant account that comes back, comes back because I happened to think of them in the shower. That isn't a process. It's one person with a good memory and a bad week ahead of them, and it fails silently — nothing alerts you to the renewal you didn't think about.

What to do about it this week

Stop asking your agent to remember, and start asking it where it wrote things down.

That's the practical version, and it changes how you set the thing up. Every time you finish a session with a conclusion — this account is worth chasing, that one is dead, this one needs a month — the useful question isn't "will it remember?" It won't. The question is whether there is anywhere outside the conversation that the conclusion now lives, and whether tomorrow's session will read it before it starts.

If the answer is no, then no amount of better prompting fixes it. You are not looking at a model problem. You are looking at a missing piece of furniture.

Why sales is next, and it isn't a trend

The preconditions arrived and nobody announced it.

The people I'm describing already have the agent. They already trust it with code, with writing, with research. They've already connected an inbox to something at least once — that's not a future behaviour, it was last month. The missing piece was never willingness and it was never capability. Nothing was holding the result.

Which is a smaller problem than it sounds and a more irritating one. Nobody is waiting for a better model. I got annoyed enough about it to go and build the missing furniture, which is a story for another week.

The desk is the harness. Sales just hasn't got one yet.

---

Next week I stop arguing and run it for real — my own inbox, calendar, CRM and Stripe, against a business with fifteen years of relationships in it — and show exactly what came back, including the parts that didn't work.