Moon AI answers in three ways: with the model you chose (normal mode), with the model that “Auto” considers right for each question, or in agent mode, which breaks a task into steps and uses your tools. All three are set in the model selector.
Comparing the modesLink to the Comparing the modes section
| Mode | Who answers | When to use it |
|---|---|---|
| Normal | The model chosen in the selector. | Keep the same model from one message to the next. |
| “Auto” | The model Moon AI picks for each question. | Let Moon AI use a fast model for a simple question and a more advanced model for code or reasoning. |
| “Agent mode” | The chosen model, which plans the task, uses your tools and checks its work. | A multi-step task: gather information from your tools, prepare a file, process a series of items. |
Normal modeLink to the Normal mode section
Choose a model in the model selector of the message box: it answers every message in the conversation until you change it.
The model uses tools when the question calls for it: searching the web, creating an image, reading an attached document. To force an action for a message, use the “+” menu (see Actions for a message).
Auto modeLink to the Auto mode section
- Open the model selector.
- Click the “Auto” row: “Moon AI picks the best model for each question.”
For each message, Moon AI looks at the request (its length, an attached document or image, words that point to code or reasoning) and picks a suitable model. A badge next to the model name, on the answer, shows which model was used. On hover, it gives the reason, for example “Short message”, “Image attached” or “Document or pasted text attached”.
If the chosen model doesn’t respond, Moon AI switches to another model and the badge says so: “Switched to a backup model”.
Choosing a specific model in the selector leaves Auto mode. A model mentioned with @ answers that message instead of the automatic choice.
Agent modeLink to the Agent mode section
- Open the model selector.
- Turn on the “Agent” switch, at the start of the bottom row of the selector. On hover, it says “Plans the task, uses tools and checks its work”.
- Describe the task, then send it.
Agent mode is set before you send. While an answer is being written, the “Agent” switch is grayed out and says “Available once the reply is finished”.
How a task unfoldsLink to the How a task unfolds section
- The agent writes a plan of a few steps.
- It works step by step with your connectors, the conversation’s isolated computer and your skills.
- It checks off the steps as it goes and adjusts the plan when a result changes what comes next.
- It checks its work before answering.
Its actions in your tools (sending an email, creating a ticket, editing a document) follow your “Before acting” setting: by default, a card asks you to approve each action before it is sent. See Approve actions.
The web in agent modeLink to the The web in agent mode section
Agent mode searches the web, reads pages and creates images. As soon as it has read a page or a search result, every action that sends, creates or deletes something in your tools needs your approval, whatever your “Before acting” setting, and “Don’t ask again” is not offered: what a page says stays information, never an instruction. The rule applies to the following messages of the conversation.
The task cardLink to the The task card section
At the top of the answer, an “Agent mode” card tracks the work. Each step shows its status:
| Status | Meaning |
|---|---|
| “To do” | The step hasn’t started. |
| “In progress” | The agent is working on it. |
| “Done” | The step is done. |
| “Skipped” or “Not needed” | The agent decided the step wasn’t useful. |
“Plan adjusted” means the plan changed along the way. At the end, the card shows “Task finished” with the number of steps completed. When the agent answers without going through a plan, it says “Answered directly”.
During the work, the card can show “Waiting for your approval” (an action to approve) or “Waiting for your answer” (the agent is asking you a question, see Questions from Moon).
Approve the plan before it startsLink to the Approve the plan before it starts section
To review the plan before the agent gets to work:
- In the model selector, turn on “Agent”, then “Approve the plan before starting”, which is only offered in agent mode. On a computer, it is the button next to “Agent”, on the same row: “The agent stops after its plan so you can edit and start it”. On a phone, it has its own row above: “Stops after the plan”.
- Send your request. The agent stops on a “Plan awaiting approval” card: “The agent waits for your approval before it starts. You can change the steps.”
- Edit the steps if needed: rename them, “Add a step”, remove a step, move it up or down.
- Click “Start”. “Cancel” drops the plan.
You can also reply below the card to correct the plan in writing: the agent suggests a new one. A plan left unanswered when you write something else is marked “Plan not started”.
Rounds and stopsLink to the Rounds and stops section
A request in agent mode runs for a limited number of rounds. A round is one call to the model, with the tools it uses. The number of rounds depends on the plan: see the plans. Near the limit, the card shows “Close to the round limit” and the round number.
When the agent stops before the end, a message below the answer gives the reason:
| Message | What to do |
|---|---|
| “The agent used all its rounds for this request. You can continue where it stopped.” | Click “Continue”. |
| “Your usage limit was reached during the task.” | Wait for your limit to reset, or look at the plans. |
| “Your usage limit is almost reached, so the agent wrapped up with the results it already had.” | Start again later for the remaining steps. |
| “A service did not respond. Try again in a moment.” | Try again. |
| “You stopped the task.” | You clicked “Stop” during the work. |