An Answer Is Free to Be Wrong, an Action Is Not

For two years the worst thing it could do
was tell you something untrue.

You read it, you doubted it, you checked.

A wrong answer cost you a minute
and taught you to read the next one harder.

Now it has hands.

It edits the file.
It runs the command.
It opens the pull request,
calls the API,
writes the row,
posts into the channel.

The arithmetic changed
the moment it stopped talking
and started doing.

A wrong answer is inert.

It sits there until you act on it,
and you are the gate.

A wrong action has already happened.

There is no reading it sceptically afterwards.
There is only the state of the world,
and a log,
if somebody thought to keep one.

Now notice what the pace does to you.

An agent works in steps,
and every step arrives
with a small green button.

The first eight are obviously fine.

By the fortieth,
your eye has learned the shape
of a step that is fine,
and it is matching shapes now,
not reading.

That is not carelessness.

That is what attention does
to a long run of nearly identical requests,
and no amount of resolve fixes it.

So change what the button means.

Separate proposing from applying.

Let it work as much as it likes
somewhere that can be thrown away.
A branch. A copy. A plan.
A dry run that prints
what it would have done.

Then read the whole result once,
as a diff,
with your full attention,
and apply it yourself.

One decision you actually made
beats forty you waved through.

And give it credentials
the size of the job.

Not yours.
Not the ones that also reach
the customer database
and the production queue
because they were already in your shell.

The question before you connect it
to anything that matters
is not whether it is usually right.

It is what happens on the day it is not,
and whether you can put that back
before anybody else finds out.

– Serguey Asael Shinder