How agents work

A coding agent is a loop: read context, decide, call a tool, read the result, repeat. Everything that makes agents useful or dangerous comes from what goes into that context.

Watch for three things in the video:

• Context acquisition. The agent reads files, runs commands, and greps. The same "trace the data" habit you are learning is exactly what a good agent does, and exactly where a bad one cuts corners.
• Tool calls are actions. A shell command that deletes a directory is not undone by the agent apologising.
• Tests as proof. When the agent says "all tests pass", ask which tests, and who wrote them.

Knowledge check

• What does an agent see when it starts on your task?
• Why is "tests pass" not the same as "the change is correct"?
• Which parts of your own workflow could an agent do faster, and which parts should stay yours?