How an agent works through a task

An agent runs a loop: read the task, choose a tool (read a file, search, edit, run a command), look at the result, decide the next step, and repeat until it judges the task done. It is powerful because it gathers its own context and checks its own work, and risky for the same reasons: it acts on its guesses, and each step builds on the last.

Knowing the loop tells you where to steer. Give it a good start: a spec, the right files and a check to run. Watch the first few steps; if it heads the wrong way, stop it early, since redirecting after two steps is cheap and after twenty is not. For multi-file changes, ask for a plan first.