Reading6 min read·Module 1

The loop, and how it recovers

Watch a good engineer fix a bug you reported. They do not type the fix immediately. First they look: open the failing file, grep for where the function is called, maybe read a test. Then they change something. Then, and this is the part that separates careful from careless, they run the test to see whether it actually worked, and if it did not, they look again. Claude Code runs exactly that cycle, and understanding it is the difference between using the tool well and being surprised by it.

The agentic loop

gather · act · verify

the failure becomes the next lap's context

gather

read + search

act

edit + run

verify

check the result

FAIL

PASS

fix login

goal verified · loop stops

A failed verify is not a dead end. Its output becomes the context the next lap gathers, so the second attempt lands where the first missed, and the loop stops only when the check comes back green.

The loop recovers: a failed verify feeds the next lap, and the cycle stops only on a passing check.

1 / 4