ESSAY · 4 MIN READ
Your agents are building last week's plan
A requirement changes on a call. The people on the call know. The agents do not. How to say a change once and have the work that rests on it wait or ask.
Phil Tran, founder · October 13, 2026
A requirement changes on a call. The people on the call know. The agents do not, so they keep building last week's plan, and you find out at review.
Think of a restaurant. The table changes its order. The waiter knows. But the ticket on the rail still says the old order, and the cook makes what the ticket says. Nobody in that kitchen did anything wrong. The cook was loyal to the ticket.
Your agents are the cook. They follow the place they were sent. If that place is old, the work is old too.
It does not even take a team. Give the same instruction in two AI sessions, in slightly different words, and two agents can ship on two versions of one decision. Now add ten people, a teammate on the other side of the world, and agents that keep working while everyone sleeps.
Why the files do not fix it
We write our rules down. CLAUDE.md, agent files, a rules folder. It helps, for a while.
Then the files drift. Nobody is paid to keep them true, and nothing breaks the day a rule goes out of date. One engineering blog calls this "rules-file rot", files "each stale in a different way, each confidently instructing your agents with last quarter's truth."
A file is a ticket someone forgot to change.
Say it once
In R II F, a decision lives in your team's record, and every job carries the decisions it rests on. So when you change one, the change goes to the work:
- Work that has not started yet on the old decision waits, or asks you first.
- A person's task resting on the old decision is flagged: "Decision changed, check your task".
- A task already done is marked "done on an old decision". You choose: keep it or redo it.
People are told, machines are held. You say the change once.
What it does not do yet
A job that is already running when the decision changes finishes today. You can see which decision it rested on. Stopping it mid-step is what we are building next.
Try it
Start your team on riif.com, put Reflex Ops on the computers that run your agents, and change one decision. Watch what waits.
Source: Unblocked blog, "rules-file rot", 8 July 2026.
Try it on your computer.
INSTALL REFLEX OPS
1. Open Terminal and paste:
$ curl -fsSL https://reflex.riif.com/install | bash2. In a new window, sign in with riif, once:
$ reflex login3. Start:
$ reflexUsing Claude Code? In your project, run reflex kit and Claude Code can hand work to Reflex Ops under a spend cap. All steps
MORE TO READ
Nothing acts on yesterday's decision
Every rule for AI says a person stays in control. At scale that breaks two ways. A story about an eighth server, and why we named it Reflex Ops.
Clicking yes is not oversight
Agents ask about everything, so we stop reading. Meanwhile the one question that matters waits all night. Answer once, and let the real question through.
Out of Claude Code usage by Wednesday?
Make Claude Code the director, not the typist. It writes the brief and checks the result; the long work runs in Reflex Ops under a dollar cap.