I keep running into teams who think their smartest AI agent should be grinding away on every single task, all day, full throttle. Then I stumbled on a post that flipped that idea on its head. The original poster, who’s built companies from zero and sold one, makes a claim that stopped me mid-scroll: your smartest AI agent should spend most of its life asleep.
I’ll be honest, I read that twice. Then it clicked, and I couldn’t stop thinking about it. So let me break down what this expert shared, myth by myth.
Myth #1: A great agent runs hot on every task
The common setup goes like this. You build an agent, then leave it running on everything. Every step, full model, full cost. According to the creator, it works beautifully right up until the invoice lands.
The truth? A frontier model shouldn’t be doing work a simple rule could handle blindfolded. Most steps in a workflow are predictable. Paying inference rates for them is like hiring a surgeon to put on a band-aid.
Myth #2: Teams burning cash are using AI wrong
This one surprised me. You’d assume the teams bleeding money are making technical mistakes. The original poster says no. They aren’t using AI wrong. They’re using it everywhere.
The fix isn’t less ambition. It’s better placement. Here’s the shape this expert recommends:
- Build agentically. Start with an agent to figure out what good looks like.
- Deploy deterministically. Harden the predictable parts into cheap, boring automation.
- Wake the AI only for edge cases that actually need judgment.
Myth #3: Every step in your workflow needs AI
Not even close, says the creator. The tell for whether a step needs AI is refreshingly simple: the answer has to depend on reading and understanding something.
In practice, that comes down to just five jobs where AI earns its keep:
- Summarising
- Analysing
- Classifying
- Drafting
- Extracting
Everything else, in the words of this savvy professional, is a rule pretending to be clever. I love that line.
Predictable step? A rule handles it, zero dollars in tokens. Needs reading and understanding? That’s the one spot AI earns its keep. Better model ships next month? Swap it in for that step, no rebuild.
Myth #4: Tokens consumed is the number to watch
Here’s the mindset shift I found most useful. The metric that actually matters isn’t tokens consumed. It’s cost per task.
The expert calls the approach spending tokens only where they genuinely earn their keep. Stop paying premium inference rates for fraction-of-a-cent work. Put the expensive brain on the hard reading-and-understanding jobs, and let cheap rules carry everything else.
The truth to act on
Build with an agent to learn the terrain. Then lock down the predictable parts into simple, cheap automation. Reserve the frontier model for the handful of moments that truly need judgment. That’s how you keep ambition high and the invoice sane.
I was genuinely impressed by how clearly the original poster laid this out. If you build AI systems, this reframing is worth sitting with. Head over to the full LinkedIn post to see the reasoning in the creator’s own words.