Stop Asking Me Halfway

Most AI agents ask too many questions.

Codex is the worst culprit. I give it a project. It gets halfway through, stops, and asks me something it could have worked out on its own.

The question is usually small. Which folder. Which version. Whether I meant this or that. The answer is almost always sitting somewhere it can already read.

So now I am the blocker. The work sits there until I see the message, switch context and reply. Then it does a bit more and stops again.

That defeats the point. I hand work to an agent so I can stop thinking about it.

On October 21 I added one rule to my agent rules file. It is one file that every agent I run reads before it starts work, including Grok Bot, Claude and Codex.

“Don’t stop a project halfway to check with Vinay. Push through all the research, setup and work you can do without asking. Pause only before an important write or delete.”

The last line matters most. I still want a stop before anything that is hard to undo. Sending a message as me. Deleting something. Changing something other people see. Those already go through a send gate where I approve the exact draft.

Everything before that point is fair game. Reading. Researching. Setting things up. Drafting. Checking its own work.

The second half of the rule is about what happens to the loose ends. Anything half-finished goes in a note. If the agent had to guess, it says what it guessed. If something needs my call, it goes on a list.

The questions come at the end, not halfway through.

That changes how a review feels. I get one finished piece of work and a short list of decisions. I can answer them in one sitting. That beats being pinged all afternoon.

It also shows me where my setup is weak. If an agent keeps guessing the same thing, the answer belongs in the rules file, so nobody has to ask again.

I prefer not to babysit my agents.