After Claude Code runs a command or edits a file, what does the result of that action do?

The principle — every action is also an observation.

Why the key is correct — Anthropic puts it directly: each tool use gives Claude new information that informs the next step. That feedback is what allows the loop to course-correct rather than commit to a plan made before anything was known.

Why the others are wrong — discarding results would leave the next step with nothing to work from. Treating one result as the verdict on the whole task ends the loop far too early. And pausing for confirmation after every action describes a different interaction model; you can interrupt at any point, but it does not stop to ask.

Remember this — a failing test is input, not a verdict.

Sources — Anthropic's How Claude Code works.