Clicking yes is not oversight

People using Claude Code approve around 93 percent of permission prompts. Ask anyone the same kind of question all day and they stop reading it. Meanwhile the one question that matters sits all night, waiting for a person who is asleep.

Both happen on the same team, in the same week.

People like to say a human stays in the loop. That sounds safe. But a loop can turn into a habit. Yes, yes, yes, and then one night the agent stops at a real decision and nobody sees it until morning.

Every answer is thrown away

The root is simple. Every answer is thrown away.

You approve the same kind of step on Monday, Tuesday and Wednesday. Each time it is a new question. Nothing you said on Monday counts on Tuesday. So the questions pile up, and the important ones look exactly like the routine ones.

The note on the fridge

When you leave your kids with a new babysitter, you write a note and stick it on the fridge. Bedtime is eight. Snacks are in the left cupboard. The neighbor has a key.

You write it once, so she does not call you at dinner. And at the bottom: call me if there is a fever.

That is the whole idea. The routine answers are written down once. The one call that matters still reaches you.

Answer once

Reflex Ops runs your team's AI agents on your own computers. When an agent reaches a call only a person can make, it stops and asks. It does not guess.

You answer once. If it is the kind of question that will come back, you make the answer a rule. The next agent with the same question does not ask. It follows the rule. The answer is kept in your team's record in R II F, with the job it was for.

So the questions you see are fewer, and the ones that reach you are the ones that need a person.

Answer once. Make it a rule. Let the next agent move on.

Install Reflex Ops

Source: The Register, 6 August 2026, reporting Anthropic's figure for Claude Code users.