I almost congratulated a prospect for something he did two years ago - while pitching him on AI

My thumb was on the send button when I checked the source.

The news was two years old.

Let me back up. I'd built an AI agent that watches my prospects overnight and tells me what to do next. That morning it had handed me a good one - a reason to reach out - and I'd already drafted the note:

“Congrats on the launch of your first store in Dubai.”

Except the store had opened two years ago. And I was about to send that to a prospect I'm trying to sell on building exactly this kind of system.

I put the phone down and just sat there for a second.

The pitch you've already seen - and why it's a trap

You know the one. Set up an AI that runs 24/7, give it a goal - “make me $1M in ARR” - and it figures out the rest while you sleep.

The mechanic underneath is real. Instead of a cron job running the same script every time, the agent wakes itself on a “heartbeat” and decides the highest-value next move. Anthropic draws the same line: a workflow follows steps you defined; an agent directs its own.

A boring routine for the 80% you can predict. The heartbeat for the 20% you can't.

The non-deterministic agent isn't the thing you hand your company to. It's a $20 specialist you call in for the edge cases - and one that's never allowed to hit send.

The night it out-thought me

One night it picked up that a company in European fish production had announced a merger.

It reasoned - on its own - that the acquired company's clients were now at risk of being deprioritised in the shuffle. So it suggested I build a small agent to find those orphaned clients before anyone else did.

I wouldn't have made that connection. A cron job certainly wouldn't.

That became a slide. The slide became a meeting. This week I sat down in person with that company's full leadership team to run an AI transformation plan.

That's the heartbeat earning its place on the 20% - spotting the wedge I'd never have seen. The economics aren't subtle:

That's what the 20% pays for.

The part nobody puts in the demo

Now back to the send button - because the same thing that lets the agent surprise you with a fish-merger wedge is the thing that nearly sent that Dubai note out under my name.

The reason the near-miss rattled me isn't the embarrassment. It's what it revealed: the agent is a bit of a black box, and it had been confidently right enough times that I'd stopped checking. I sell building these things to clients. So the one time it makes me look like I didn't do my homework isn't a cosmetic problem. It's the whole business.

So I coached it, the way you'd coach a junior who'd just made a confident mistake. I sent a voice note on Telegram:

Review the article yourself, attach the source URL whenever you suggest I reach out, and filter out any signal older than two weeks.

It rewrote its own skill file. I never touched a config. Now every draft ships with its sources, and two-year-old news doesn't reach me.

That's the shape of the whole system - the heartbeat thinks, but it only ever drafts; I decide. Drafts-only isn't a rule I bolted on. It is what determinism looks like at the moment of action:

• It drafts, never sends. Nothing it produces touches the outside world without me.
• Every draft ships with its sources - verify a signal in five seconds.
• An LLM-as-judge screens drafts for ICP fit - the client, the project, the angle - before they reach me.
• I coach it in plain language, and it edits its own skills.

And sometimes the honest call is to take a job away from the heartbeat. I had one scraping derivatives and settlement data off a regulator's website. It kept inventing new paths to scrape, the output was too messy to trust - and it burned through twice my Codex usage in under a week. So I demoted it to a cron job: the same checks every day, same comparison.

Boring. Reliable. Correct.

Build the boring 80% first

If you want one of these, start tomorrow, small:

• Point one agent at your open deals - one job you understand well enough to catch when it's wrong. Not your whole company.
• Run it once a day on a $20 sub. A flat subscription can't hand you the surprise bill a metered API can - and once a day is plenty. I learned “once” the hard way: the scraper that ran constantly is the one that doubled my Codex bill in a week.
• Let it draft the next move - with sources - and approve every send yourself.
• Keep determinism for the 80%. Reach for the heartbeat only on the edge you couldn't have scripted: the merger, the conference talk, the thing a cron would miss.

The promise online is that you set a goal and the agent runs your company. That's not the version that works. The version that works is more useful and less sexy: a $20 specialist that handles the part you can't script, drafts instead of acts, and shows its work.

Here's the part that took me a near-miss to understand: you don't keep it on a leash because it's bad.

You keep it on a leash because it gets good enough that you stop checking - right up until the moment it asks you to congratulate a stranger for something he did two years ago.