How it works
Approvals
What the approval dialog shows, what each choice means, what happens when you deny, and which steps always stay with you.
The approval dialog
Before a sensitive action, ActionWright stops and shows a dialog titled Action Approval Required. The action does not run until you answer. The same dialog is used for the file agent, the browser agent and Computer Use.
Which actions ask first is listed on each agent's page:
- Files: What it can do, and what asks first.
- Browser: What asks first.
- Desktop apps: Computer Use.
What the dialog shows
- The action, at the top, for example Move Items, Fill Form Field or Click on screen.
- A preview of what will happen. For a file action: how many files and folders are affected, a one-line effect, and every item with its full path. For a browser or Computer Use action: that one action, described in rows, such as the site or app, the button, the recipients, and the text that will be sent.
- A line under the preview. When Always Allow is offered, it reads "You can click "Always Allow" to skip this dialog for this action type in future sessions." Otherwise it says why the action needs you, for example "This action always requires explicit approval." or "The effect of this action could not be determined."
- The buttons, left to right: Deny, Always Allow (only when offered) and Allow Once.
The dialog stays on top of other windows. If the task window is not in front, Windows also shows a Confirmation required notification. If you started the task by voice, ActionWright says "Confirmation required" and the name of the action.
The preview does not always show everything. For example, a file preview does not show the text a file will get or the destination of a move. See What the preview shows.
Your choices
| Choice | What happens |
|---|---|
| Allow Once | This one action runs. The next action of the same kind asks again. |
| Always Allow | This action runs, and ActionWright stops asking for this type of action: in every folder, in later tasks, and after a restart. Offered only for some file actions. See Always Allow. |
| Deny | The action does not run. See When you deny. |
esc, or closing the dialog | The same as Deny. |
Always Allow
- It is offered for these file actions: Create Folder, Write File, Edit File, Rename Item, Move Items and Copy Items.
- It is never offered for actions that always ask, such as Move to Recycle Bin.
- It is never offered for browser or Computer Use actions. Each of those approvals covers one action only.
- It applies to the whole action type, not to one folder or one file. After Always Allow for Write File, the file agent can replace the text of any file in any folder you choose, in later tasks too, without a dialog.
- ActionWright remembers the choice for your Windows user, in its own data on your PC.
When you deny
- The action does not run.
- The agent's current piece of work ends. Other actions it had lined up in the same step are cancelled too.
- Changes made before the denied action stay as they are. Denying does not undo them.
- The task goes on. ActionWright learns that you cancelled the action, and may try another way, possibly with another agent, ask you what to do, or stop and report what it did.
- In the browser, an action you denied is not done again in that chat unless you ask for it again. See Protections.
If you want the whole task to end, click Deny, then click Stop in the task window.
A waiting approval
- An approval has no time limit. The dialog waits until you answer it.
- Nothing else in the task happens while it waits. To end the task instead, see When you deny.
Browser approvals
- The header names the kind of action, for example Browser Action, Fill Form Field, Upload File, Navigate or Browser Batch.
- The preview shows what the action will do: the site, the button or field, the recipients, the text that will be sent and the form fields, plus a line on why it asks. Values of protected fields, such as passwords, are not shown.
- A download that could run a program, such as an installer, asks before it is kept. The preview shows the file's name, type and size, and the site it came from. If you deny it, the file is not kept.
For the full list, see What happens with each kind of action.
Computer Use approvals
- Before some actions on your screen, Computer Use asks you first, one action at a time.
- The header names the action, for example Click on screen, Type text or Press keyboard shortcut.
- The preview describes the action and names the app window it is in, plus a line on why it asks.
Which actions ask, and what Computer Use does on its own, is on Computer Use.
How ActionWright decides what may run
Every task involves three parties: you decide what is allowed, the agent does the work, and a separate check makes sure each consequential action matches what you allowed. This applies to the browser agent and Computer Use. The file agent does not use this check: its changes ask first unless you chose Always Allow. See Always Allow.
- Your words set the permissions. When you send a task, a separate step reads only your own words, your request and your answers to its questions, and works out which kinds of consequential actions this task may take. It never reads web pages, files, the screen or the agent's work.
- The agent only proposes. The agent doing the work never decides what it is allowed to do. This way, outside content such as a web page, a document or an email cannot widen what is allowed: even if it asks the agent to do something, it cannot add a permission. Only your words can.
- Each consequential action is checked before it runs. A separate check looks at the action itself, what it would do and where, without the agent's conversation or reasoning. Then fixed rules decide what happens:
- It runs if your request allowed that kind of action and the action serves your task.
- You are asked if your request did not allow that kind of action, or if the check can't tell whether it serves your task.
- The agent looks again if the action's effect isn't clear, and you are asked if needed.
- It doesn't run, and no dialog appears, if it clearly has nothing to do with your task. The agent is told, and continues another way.
- It isn't done again if the same action was already done, or you denied it, in this chat.
- Paying or placing an order is built to always ask.
What happens with each kind of action
| Kind of action | What happens |
|---|---|
| Paying, buying, placing an order or moving money | Always asks |
| Sending a message or email, posting, or submitting a form | Runs if your request asked for it, otherwise asks |
| Deleting something | Runs if your request asked for it, otherwise asks |
| Changing settings, accounts or permissions | Runs if your request asked for it, otherwise asks |
| Accepting terms or agreements | Runs if your request asked for it, otherwise asks |
| Installing or removing software, or restarting the PC (Computer Use only) | Runs if your request asked for it, otherwise asks |
| Uploading a screenshot the agent took (browser only) | Always asks |
| Downloading a file that could run a program (browser only) | Always asks |
| An action whose effect isn't clear | The agent looks again; asks you if needed |
| An action clearly unrelated to your request | Doesn't run, no dialog; the agent is told and continues another way |
| Reading, finding your way around a site, ordinary work in an app, or uploading a file you attached | Runs without asking |
These checks also run at your AI provider. They are a safeguard, not a guarantee: what is on a page or the screen can still make an action look different from what it is. That is why ActionWright is built to ask before paying, and why actions with an unclear effect ask you. Read each approval before you allow it.
Steps that stay with you
Some steps are not done for you: signing in, CAPTCHAs and other robot checks, two-factor codes, and entering payment details. The browser agent does not type passwords, one-time codes, card or bank details, or ID numbers. Computer Use is built to hand these steps to you too, so stay nearby when a task may reach a sign-in screen in a desktop app.
When a task reaches such a step, the agent stops and hands control to you:
- Under the result in the task window, a message says what you need to do. If the window is not in front, Windows shows an Action required notification.
- Do the step yourself, in the agent browser or on your screen. Click Show Agent Browser if the agent browser is not in front.
- Optionally, type a short note in Reply (optional)..., for example "I chose the second address".
- Click Done, or press
enterin the reply box.
The agent then looks at the page or the screen again and continues. A hand-off has no time limit. To end the task instead, click Stop.
Tips
- Read the preview. It shows what will change, or what will be sent and to whom.
- Deny when you are unsure. A denied action does not run, and you can ask again in clearer words.
- Be specific in your request. A clear request makes the preview easier to judge, and helps the check block what you did not ask for.
