ARTICLE · CONTROLS FOR AI WORK · PART 1 OF 6 · State programming

Opinion or execution?

A machine always knows which mode it’s in: running, stopped, homing. A conversation with an AI agent doesn’t. You ask for advice, then ask for something exact, and it’s still giving advice. You get picky, then ask an open question, and it’s still picky. Most of the trouble people have with AI starts right there: the AI agent is still doing the last job. The fix is old and simple. Know which job you want, and say so.

The same questions apply to a sales pipeline, a purchase order or a month-end close: what stays fixed, what needs judgment, and who decides?

What the AI field has found

A team of AI researchers tested top AI models on the same tasks two ways: everything given at once, or spread over a conversation the way people really talk. In conversation, performance dropped by an average of 39% across six kinds of tasks. The models weren’t much less capable; they were much less reliable. They made assumptions early, then leaned on those assumptions for the rest of the conversation, and had trouble recovering when they’d guessed wrong. (Laban and others, “LLMs Get Lost in Multi-Turn Conversation,” ICLR 2026.)

That matches what people see every day. A long conversation builds up its own momentum. The AI agent settles into whatever it decided the job was, and it keeps doing that job.

From the shop floor: one mode at a time

On a machine, running, stopping and homing are different jobs. A state-based program gives each one its own section, and only the active section runs. A homing routine should never take over because some unrelated line of logic happened to come true. Moving from one mode to the next is a deliberate step, with its own conditions, never an accident.

The reason is simple: when code from the wrong mode runs, you don’t really know what the machine will do.

AI has two modes that matter most, and they need to be run differently:

  • Opinion. You want its judgment. Give it room. Tell it the whole story, the messy version, and let it bring up something you didn’t think of. Hold it too tight here and you only get back the answer you already had in mind.
  • Execution. You want it to follow the steps exactly: enter the fields, format the file, run the procedure. No suggestions, no improvements, no “you might also consider.”

Both are useful. Mixing them up is where it goes wrong.

Opinion mode: talk, don’t type

When I want advice from an AI agent, I don’t type fifteen careful words. I talk for five minutes. What’s going on, what I’ve tried, what worries me, what the people involved are like. Then I ask what it thinks, and I argue with it when the answer comes back generic.

That’s how you’d brief a sharp new engineer out of school. You wouldn’t give them three minutes and walk away. They have the textbook half memorized, so they’re very good at the standard answer. What they don’t have is the judgment of when: the nuances, the edge cases, the things nobody wrote down. You bring that.

There’s one more thing AI does in opinion mode that a person would recognize. It tries to solve the problem. Sometimes the person in front of you doesn’t want a solution; they want to be heard. That’s a time to leave the AI agent out of it.

Execution mode: say exactly what, and nothing else

Execution is the opposite. The AI agent should follow the procedure the way a machine follows its program. A comma out of place can stop a printer. So the request says exactly what to do, and says so up front: “Do exactly this. Don’t add suggestions.”

In an AI agent, this is where the procedures and skills come in (Part 2). The AI agent doesn’t have to guess the mode, because the procedure says which job this is and what done looks like.

Changing modes on purpose

The hard part is the switch. On a machine, a mode change is an explicit transition. In a conversation, it’s easy to drift from one to the other without noticing.

The A.N. Creation Office’s Project Manager Bot has a rule for exactly this. It’s one of its principles:

Principle: Re-anchor. The Project Manager Bot must start every interaction by asking how the current topic fits into the project. It must ask this before answering.

Before it answers, it places the new topic: is this part of the project, a side question, a new idea? That’s a mode check, the same as a program checking its state before it runs any code.

A long conversation can talk itself into a corner

There’s one more habit I use, and it comes from the same research finding. After a long working session, the AI agent has often formed an opinion. The conversation has built up enough that it has convinced itself. And when everything it says starts agreeing with you, that’s a warning sign, not a good one.

So at the end of a long session, I take the conclusion, just the conclusion, to a fresh session with a clear context. Sometimes the fresh one says, in effect, “why would you do that?” I’ve seen an AI agent disagree with itself more than once. For anything that matters, I’ll also run it past a second model.

It’s the same instinct as not letting one long meeting be the only vote. Sleep on it, or ask someone who wasn’t in the room.

Try this

  • Start your next request by naming the mode: “I want your opinion” or “Do exactly this.”
  • In opinion mode, talk for five minutes instead of typing a sentence. Argue with the first generic answer.
  • When you switch modes in the same conversation, say so.
  • After a long session, paste only the conclusion into a fresh chat and ask what’s wrong with it.

Sources

These pages sit with the rest of the Advance Now family. The hotbotspot tool library writes the methods up for AI. A.N. Creation Office is where the public AI agents run as open experiments.