AI agents: when the model carries out the steps itself
An AI agent is a model that does not just answer but also acts: it reads files, runs commands, fixes code and checks the effect, moving step by step toward a goal.
Chat versus agent
- Chat: you ask, the model answers, and you handle the copying and running yourself.
- Agent: it is given a goal and has access to tools, so it carries out a series of actions on its own.
What an agent consists of
- A model that plans and makes decisions.
- Tools: reading and writing files, a terminal, search, a browser.
- A loop: action, observing the result, the next decision, until the task is finished.
- Working memory, that is, the context that holds the plan and the steps so far.
What they are good for
- Tasks with a clear success criterion, for example "make the tests pass".
- Repetitive changes across many files.
- Tidying up and documenting a project.
- A first investigation of an error and a proposed fix.
Where it gets risky
- An agent can make many changes at once, so they are harder to judge.
- A mistake at an early step can carry over to the following ones.
- Access to the terminal and files means the ability to delete or modify something important.
- An agent may carry out commands hidden in the content it processes if you do not filter it.
Common-sense rules
- Grant minimal permissions and work on a copy of the project.
- Save the state in Git before every larger task.
- Require a plan before action and confirmation before irreversible operations.
- Give small, well-described tasks instead of "build the whole app".
- Review the diff when the work is done.
An overview of tool categories can be found in the text on AI programming tools.
