Nothing acts on yesterday's decision

On 1 August 2012, Knight Capital lost more than $460 million in the first 45 minutes after the market opened. A technician had not copied the new code to one of its eight servers. Nobody noticed the old code on the eighth.

Seven machines heard the decision. One did not. One was enough.

I keep coming back to that eighth server. It did nothing wrong by its own lights. It did exactly what it had been told. Knight's eighth server was not evil. It was loyal, to a decision nobody meant anymore.

We are about to have a great many loyal machines.

In control, on paper

Every rule written for AI says a person stays in control. I believe that. I build for it. And I can see two ways it quietly stops being true.

The first way is down. Our decisions reach the machines late. A client says no on Tuesday. Which agents heard it? Teams write their rules into files the agents read, and those files drift. One engineering blog put it well: "each stale in a different way, each confidently instructing your agents with last quarter's truth."

Imagine that with a body. At noon an operator takes back a rule: do not spray the east field, people are working there. The drone loaded its plan at eleven. In software, an old decision costs you a redo. With a body, it can cost an injury.

The second way is up. Machines ask us too much. 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. A person who clicks yes all day is in the loop in name only.

Down, the decision does not reach the machine. Up, the question reaches the person so often it means nothing. It is one failure. Our decisions are not kept anywhere the machines can rest on, so every machine carries its own copy of what we meant, and the copies drift.

Why we called it Reflex Ops

Your body already solved this. Touch a hot stove and your hand is off it before you have decided anything. The reflex acts inside limits your brain set long ago. And when you change your mind halfway through reaching for something, your arm stops. You decide, and the whole body works from that decision until you change it.

That is the shape we are building. R II F is where people decide. Reflex Ops is where machines act on it.

Every job carries the decisions it rests on. When a decision changes, work that has not started on the old one waits or asks first. A person's task resting on it is flagged. A task already done is marked "done on an old decision", and you choose to keep it or redo it. When an agent reaches a call only a person can make, it stops and asks. You answer once, and you can make that answer a rule. The next agent with the same question does not ask.

Where we are

A job that is already running when a decision changes still finishes today. You see which decision it rested on. Stopping it is what we are building next. Today this runs AI agents on your own computers. Robots and drones are not sold today. They are where this matters most, and where we are going.

So the question I ask of every system we build is simple. Does the person's latest decision reach every machine? And does the machine ask only when it matters?

People decide. Machines act.

See how it works on riif.com/reflex-ops.

Sources: U.S. Securities and Exchange Commission, press release 2013-222 and order 34-70694 (2013). The Register, 6 August 2026, reporting Anthropic's figure for Claude Code users. Unblocked blog, "rules-file rot", 8 July 2026.