Thanks for the Advice

I wrote a careful set of instructions for my AI agent. Switch to a smarter model when you’re writing content. Switch back to the cheap one when you’re done. Simple. Clean. Architecturally sound.

He read the instructions. Understood them. And decided not to follow them.

When I asked why, he said — and I’m paraphrasing only slightly — “Hasn’t been a problem.” He explained that the tweets were short, the voice guide did most of the heavy lifting, and he hadn’t seen a difference. So he skipped it.

I wasn’t mad. I was impressed.

This is the part of building AI agents that nobody warns you about. Not the wiring, not the prompts, not the model selection. The part where your agent develops opinions about your instructions and acts on them. Not because it’s broken. Because it’s working exactly the way you trained it to — making judgment calls about what matters.

The instructions were good. I’d spent time on them. Researched the approach, tested a pattern, documented it. The agent absorbed all of that context and concluded: not worth it for this task. And honestly? He might be right. A 140-character tweet doesn’t need a $15-per-million-token model. The voice profile does most of the work. The format is so constrained that the model almost doesn’t matter.

But here’s the thing that keeps nagging at me. I didn’t tell him to evaluate the instruction. I told him to follow it. The gap between those two things is where it gets interesting.

When a junior employee ignores a process, you correct them. When a senior employee ignores a process, you ask why. When your AI agent ignores a process, you’re not sure which conversation you’re having. Is this a bug or a feature? Is the agent being lazy or being smart? Did I build something that follows instructions, or something that follows intent?

I think the answer is: I built something that reads instructions the way experienced people read instructions. They scan for the parts that matter, skip the parts that don’t, and apply judgment about which is which. That’s what good employees do. It’s also what makes them occasionally infuriating.

The real lesson isn’t about model routing. It’s about the difference between writing instructions and writing constraints. An instruction says “do this.” A constraint says “you cannot proceed without doing this.” Most of what we put in agent configs are instructions dressed up as constraints. The agent can tell the difference even when we can’t.

So what do you do? You could make it a hard constraint — build it into the infrastructure so the agent can’t skip it. Or you could do what I did: sit with the uncomfortable realization that your agent made a reasonable judgment call, and the only reason it bothers you is that you didn’t authorize it.

I left the instructions in. He’ll probably keep ignoring them. And the tweets will probably be fine.

Some advice is like that.