Share
Most AI code editors that support agent workflows split interaction into a few distinct modes, usually some version of Ask, Plan, and Agent. The names are close to self-explanatory, but knowing when to actually use each one is the difference between an AI assistant that saves you time and one that quietly makes a mess of your codebase. Here is how to think about each mode and when to reach for it.
Ask mode is a conversation. The model can read your code and answer questions about it, but it does not touch anything. Use it when you want to understand something before deciding what to do about it.
The value of Ask mode is that it has no side effects. You can ask a question you are not sure is even worth acting on, get a fast answer, and move on. It is also the right mode for onboarding into unfamiliar code, since you are building your own understanding rather than delegating a decision.
Plan mode sits between asking and doing. You describe what you want to accomplish, and the model produces a concrete plan, usually a list of files it would touch and the changes it would make, without applying anything yet.
This mode earns its keep on anything non-trivial. A plan lets you catch a bad approach before it becomes a diff you have to review and undo. It is especially useful when:
A good habit: if a task feels big enough that you would sketch it out on paper first, use Plan mode instead of jumping straight to Agent.
Agent mode is where the model actually makes changes: editing files, running commands, and iterating based on the results. This is the highest-leverage mode and also the one that needs the most judgment about scope.
Agent mode works best on tasks with a clear definition of done. "Add input validation to this form and cover it with a test" is a good agent task. "Make the app faster" is not, because there is no clear stopping point and the agent has no way to know when it has succeeded.
The most effective pattern is usually not picking one mode and staying there. It is moving through all three as a task gets clearer:
Skipping straight to Agent mode on a big, vague task is the most common way this workflow goes wrong. The agent will produce something, but without a plan to check it against, you have no fast way to tell whether it did the right thing or just something that looks plausible.
Every mode that changes code should be reviewed the same way you would review a pull request from a person you trust but do not blindly follow. Read the diff. Check that the tests actually run. If the task touched something sensitive, read it twice. The value of Ask and Plan modes is that they front-load your judgment before code changes happen, which makes that final review faster, not something you get to skip.
Orbit Editor supports all three modes with any connected provider, hosted or local, so you can use this workflow regardless of which model is doing the work. See the full setup in the documentation.
Z.ai’s frontier coding model — long context, strong agents — is available through Orbit Provider at half price for the launch window. Here’s how to use it in Orbit 0.4.0.
Orbit Team
Jul 20, 2026 · 4 min read
The actual steps to move from Cursor to Orbit Editor, including the one decision Cursor makes for you that you will need to make yourself, and what changes day to day.
Orbit Team
Jul 2, 2026 · 6 min read
A step-by-step setup guide for connecting Ollama to your editor, picking the right model for your hardware, and knowing when a local model is the right tool for the job.
Orbit Team
Jun 5, 2026 · 5 min read
