My boss wants me to experiment with an AI agent for development, but our company is small and has a limited budget. So far, I've only used a free AI account for asking about features, functions, standards, and common industry practices. I haven't had it generate code for me.
I'm open to trying an agent, but I don't want it creating entire features or dumping huge blocks of code into the project. I'd prefer to ask questions most of the time and occasionally have it make small, specific changes. Ideally, I'd be able to review exactly what it wants to modify, compare the changes with a diff similar to Git, and accept or reject them individually.
Do tools support this kind of workflow, or do AI agents generally require giving them broad control? I'm mainly looking for something that understands the project context while allowing me to remain responsible for reviewing and approving everything.
3 Answers
Visual Studio Code’s built-in chat supports this pretty well. You can use it mainly for questions, or switch to an agent-style mode when you want it to edit files. When it proposes changes, you can inspect the diff and accept or reject individual lines or edits, much like reviewing a Git change.
Tools such as Claude Code integrated with JetBrains IDEs offer both a proposal mode and a more permissive editing mode. In the normal mode, changes appear in a side-by-side diff for approval. You can also tell the agent to make only small, targeted edits, and put those preferences in a project instruction file so they apply to future sessions. Similar ask-versus-agent modes are available in products such as GitHub Copilot across VS Code, Visual Studio, and JetBrains IDEs.
Yes, this workflow is very common. IDE integrations such as those in JetBrains products can answer questions, suggest code, or make changes while still leaving you in control. Proposed edits are shown as diffs, and you can review everything before accepting it. The important part is knowing when the output is unreliable and reviewing every generated change rather than treating the agent as an authority.
A good starting point is to keep the tool in question or proposal mode, ask it to explain its plan first, and commit your work before approving any edits. That gives you an easy rollback path without surrendering control.