You have watched agents work for a whole module now. Sometimes they go wrong. Here is what
to do, and how to choose.

Response 1: Stop it

Use this the moment you see it doing something you did not ask for.

You do not need a reason beyond “that is not what I asked”. Stopping costs nothing. Letting
it finish costs you a large change to review.

Signs to stop immediately:

  • It is editing a file you did not mention.
  • It is installing something.
  • The plan has grown from one thing to six.
  • It is changing a test instead of the code the test is testing.

That last one deserves its own warning. When an agent cannot make a test pass, changing the
test is a very tempting move, and it produces a green result that means nothing.

Response 2: Correct it

Use this when it is close but has misunderstood one thing.

Correct it with information, not volume. Saying “no, that is wrong” again, louder, does
very little. Saying what you actually want does a lot:

Bad:   That is not right, fix it.
Good:  nextGap(1, "easy") returned 3. It should return 2, because 2.5
       rounded down is 2. The rounding is the problem, not the multiply.

The second one tells it what you saw, what you expected, and where to look.

Response 3: Start again

This is the one people use too late.

If you have gone round the loop three times and it is still wrong, stop correcting. Each
correction is now sitting in the context window, along with three wrong attempts. The agent
is reading its own confusion. Adding a fourth message makes that worse, not better.

git restore .        # throw away this attempt entirely

Then start a fresh session, and write a better request — one that includes what you learned
from the failed attempt.

That last part is the whole point. You did not waste the three attempts. You bought
information about what was unclear. Put it in the new request.

Choosing between them

What you see Do this
Doing something you did not ask for Stop
Right idea, one detail wrong Correct
Three failed attempts at the same thing Start again
You have stopped reading its output carefully Stop, take a break

That last row is honest advice. Approval fatigue is real. When you notice you are clicking
yes without reading, that is the most dangerous state in this whole course, and the fix is
not willpower. It is a break.

Why starting again feels harder than it is

Throwing away twenty minutes of the agent’s work feels wasteful. It is not, for a simple
reason: the agent will redo it in about ninety seconds.

The expensive thing in this process is your attention, not its output. Protect the
expensive thing.

Try this before the next lesson

  1. Deliberately give a vague request, let it go wrong, and practise git restore ..
  2. Write a correction using the good pattern: what you saw, what you expected, where to look.
  3. Think back over this module. How many times did you continue when you should have
    restarted?
PreviousNext