The agent loop
The assistant works in a loop known as ReAct: think, act, look at the result, repeat. Each pass can read your document, change it, or both.
One pass
- The assistant builds a prompt from your request and the document context.
- It sends the prompt, with the list of available tools, to the model.
- If the model asks for a tool, the assistant runs it. A tool call that fails is tried up to three times in total. A call that takes longer than 120 seconds is stopped and not tried again.
- The result goes back to the model, and the loop repeats.
- The loop ends when the model gives a final answer. Five steps before the step limit, the assistant tells the model to wrap up. If a Quick or Standard request reaches its limit, it moves up to the next size and keeps going. If a Deep request reaches its limit, the assistant asks whether to continue, wrap up with what it has, or cancel.
Request sizes and step limits
The assistant sorts each request into a size and sets a step limit to match.
| Request size | Step limit |
|---|---|
| Direct (handled without the model, for undo and redo) | 0 |
| Quick | 12 |
| Standard | 20 |
| Deep | 100 |
After its answer, the assistant may check and polish its own work: at most 1 extra pass for a Quick request, 2 for Standard and 15 for Deep. A new message from you is classified afresh.
Tool risk levels
Every tool is labelled low, medium or high risk. A tool also records whether it changes the document and whether it has effects outside the document. The full list is on Agent tools.
When you are asked to approve
You choose one of two modes.
- Ask before edits. The assistant asks before any tool that changes the document and before any high-risk tool.
- Accept all edits. The assistant runs those tools without asking.
Some actions ask in both modes:
- Each kind of export asks before its first write to a folder in a request; further exports of the same kind to the same folder during that request do not ask again.
- Scripts always ask.
Edits in Writer that are recorded as tracked changes do not ask, because you can review and reject them. Tools that add comments still ask.
If nobody can answer the prompt, the action is refused. The assistant does not guess.
Denials
If you deny three actions in a row, the assistant ends its turn. After 20 denials in total, it stops asking and blocks those actions.