How it works
How ActionWright works
The parts of ActionWright, how a task is planned and routed, what limits each run, and what ActionWright remembers.
The parts of ActionWright
ActionWright is one app on your PC, made of several parts. You talk to the app. A coordinator called the orchestrator plans the work and hands each piece to an agent. The AI model that does the thinking runs at your AI provider.
| Part | Runs on | What it does |
|---|---|---|
| The app | Your PC | The task window, voice, the tray icon, Settings and the approval dialogs. This is where you give tasks, answer questions and approve actions. See Task window and Voice. |
| The orchestrator | Your PC | Reads your request, plans the steps, searches the web and reads pages for simple lookups, asks you when it needs more detail, hands work to the agents, and writes the answer. |
| File agent | Your PC | Works with files and folders, only inside the folder you choose and on files you attach. See Files. |
| Browser agent | Your PC | Uses websites in its own Chrome window and profile. See Browser. |
| Computer Use | Your PC | Looks at your screen and uses desktop apps with the mouse and keyboard. It is on by default. See Computer Use. |
| Your AI provider | The provider's servers | Runs the AI model that every part above uses to decide what to do next. All parts use the provider, key and model you set in Settings. See AI providers. |
The agents are separate from each other and from the orchestrator, and each one has only the tools for its own job. The file agent has no browser. The browser agent cannot open files on your PC, and the only files it can upload from your PC are the ones you attached. They run on your PC as part of ActionWright.
When Computer Use is off in Settings, the orchestrator cannot use it. The other agents keep working.
How a request flows
- You ask. You type a request in the task window, or say it in Voice mode. Files you attach and the root folder you choose go with it.
- The orchestrator plans. It reads your request and the conversation so far, and decides on the next step. It decides again after every step, so the plan can change as it learns more.
- It looks things up. For a simple question, the orchestrator searches the web and reads a few pages itself, without opening a browser. See Where your data goes for the search service it uses.
- It hands work to an agent. For real work on files, websites or desktop apps, the orchestrator gives that piece to the file agent, the browser agent or Computer Use. The agent works through it step by step and reports the result back to the orchestrator. A task can use several agents, one after another. Each step appears in the Execution Trace.
- It asks you when it needs you. If something is missing, the orchestrator asks a question in the task window. Before a sensitive action, the agent waits for your approval. For a sign-in, a CAPTCHA or a similar step, the agent hands control to you until you click Done. See Approvals.
- It reports back. The answer appears under Result. In Voice mode, ActionWright also speaks it.
- You stop or continue. You can stop a task at any time with Stop or
ctrl+esc. Your next message continues the same conversation, and ActionWright can pick up work you stopped. See Stop a task and Continue a stopped task.
What the AI model sees
To decide each step, the AI model has to see what ActionWright sees. All of this is sent to your AI provider:
- Your messages, and the conversation so far.
- The text and pictures of files you attach.
- The path of your root folder. When a file task starts, the names and sizes of the files and folders in your root folder and one level below it. Later, the names and sizes of the files the file agent lists or finds.
- The content of files the agents read.
- Search results, and the text of web pages the orchestrator or the browser agent reads.
- Screenshots of the agent browser, and of your screen while Computer Use works, with the names of open windows.
- Your answers to its questions, and the result of every action.
ActionWright hides the values of password fields, and of some other sensitive fields, before page content and screenshots are sent. It cannot recognize every sensitive detail on a page or a screen.
With voice, your recordings and the text of spoken answers go to your voice provider. For every destination, and what your provider may do with the data, see Privacy.
Limits per run
Every task has step limits, so a task that goes in circles cannot run forever. A step is one action, such as a web search, reading a page, a click, or reading a file.
| Who | Step limit | How it counts |
|---|---|---|
| The orchestrator | 100 | Its own steps, such as a web search, reading a page or handing work to an agent. The count starts again for each message you send. |
| File agent | 300 | Per run: each time the orchestrator hands it work, it gets the full limit. |
| Browser agent | 200 | Per run. |
| Computer Use | 500 | Per run. |
You cannot change these limits in Settings.
There is no time limit on a task or on an agent's run. A long task keeps going until it is done, reaches a step limit, or you stop it. Only a single request to your AI provider has a waiting limit. If your provider does not answer in time, ActionWright retries, and then pauses the task as a provider problem. See AI provider errors.
When a limit is reached
- The task pauses. The task window shows "Step budget reached" with the steps used, and the buttons Yes, Continue and No, Stop. If the task window is not in front, Windows shows a Step limit reached notification.
- Yes, Continue gives the orchestrator or the agent that ran out a fresh set of steps, as many as its limit. The task goes on where it stopped. If it runs out again, you are asked again.
- No, Stop ends the task. ActionWright asks your provider for a short summary of what was done and what is left, and shows it under Result.
- The pause has no time limit. It waits until you choose, or until you click Stop or New Chat.
If a task keeps reaching a limit, read the Execution Trace before you continue. The task may be stuck on something you can fix, or it may work better split into smaller tasks.
What ActionWright remembers
ActionWright's memory is built for two goals: keep the AI focused on what matters right now, and keep your provider costs down. Every step of a task is sent to your AI provider, so a record that only grew would get slower, more expensive and easier to lose track of. ActionWright keeps what is current in full and shortens the rest.
While a task runs
ActionWright keeps a working record of every step: what it searched for, the pages and files it read, the screenshots it took, and the result of each action. It sends this record to your AI provider each time it decides on the next step.
Each agent keeps its own record
The orchestrator and each agent work from separate records. When the orchestrator hands work to the file agent, the browser agent or Computer Use, that agent starts a fresh record that holds only its own task. Every click, page and file it handles stays in that record. When it finishes, it hands the orchestrator only a summary of the result. The browser agent's summary also says which results it could confirm and which it could not.
This way the orchestrator stays focused on your request instead of on every click, and each agent stays focused on its own piece of the work.
Long tasks are shortened
As a record grows, ActionWright shortens it automatically. Fixed rules in the app do this, not the AI, so nothing is reworded or invented:
- Your messages and the AI's decisions are not changed. Only the results of actions are shortened.
- Recent results stay in full. How many stay depends on the kind of result. A page read or a screenshot goes out of date quickly, so only the latest ones stay in full. Results of file and browser actions stay longer.
- Older results become a short note with their outcome and a brief summary. For page text and file text, short samples from the beginning, the middle and the end are kept, in case they are needed again.
- Older screenshots are removed.
- On very long tasks, the oldest steps are folded into one compact list of what was done and how each step went. The newest steps stay in full.
If a long task seems to forget an early detail, repeat it in your next message.
After a task ends
When a task ends, ActionWright keeps only a short version of it for follow-ups in the same chat:
| What | Kept for follow-ups? |
|---|---|
| Your messages, including the text and pictures of files you attached to them | Kept |
| ActionWright's answers to your messages | Kept |
| Answers you typed to its questions during the task, and files you attached to those answers | Dropped |
| Search results and the text of pages it read | Dropped |
| Contents of files the agents read | Dropped |
| Screenshots | Dropped |
| The intermediate steps and the result of each action | Dropped |
Each follow-up sends what is kept to your provider again, including the text of the files you attached. Long chats therefore cost more. Start a New Chat when you move to another topic.
Research and comparisons
Because the pages a task read are dropped when it ends, the answer is what carries the findings forward. When you ask ActionWright to research or compare, for example "compare these three laptops", it aims to put the key details from its sources, such as prices, dates and specifications, into its answer, so your follow-up questions can use them. For a quick lookup it keeps the answer short.
Stopped tasks
- A task you stop stays in the conversation: your request is kept, with the note that the task was stopped. This also applies to a paused task that you end with Stop.
- If an agent was working when you stopped, ActionWright keeps that agent's working record in memory, so you can ask it to continue in the same chat. See Continue a stopped task.
- Only the most recent stopped runs of a chat are kept. When a new one is kept, the oldest is forgotten, and New Chat forgets them all.
Where it is kept
- The conversation is kept in memory on your PC only. ActionWright does not save it as chat history, and there is no history to reopen.
- New Chat starts a conversation without the earlier one. ActionWright does not use the earlier conversation again.
- Quitting ActionWright clears all conversations.
- ActionWright keeps a local activity log of what it did, which can include parts of your requests. See What stays on your PC.
- Your AI provider receives the conversation with every request. What it keeps is governed by its own terms. See Your provider's terms.
Troubleshooting
| Problem | Likely cause | What to try |
|---|---|---|
| A long task forgets an early detail | Older steps were shortened. | Repeat the detail in your next message. |
| A follow-up does not know a price or fact from the last task | Only the final answer is kept, not the pages or files the task read. | Ask again. Next time, ask for the detail in the answer. |
| A follow-up does not know an answer you gave during the last task | Answers to questions are dropped when the task ends. | Include the answer in your new message. |
| "Step budget reached" appears again and again | The task is stuck, or too big for one run. | Read the trace, then click No, Stop and split the task into smaller ones. |
| A task waits and nothing happens | It is waiting for your answer, approval or Done. | Look under Result in the task window and for an approval dialog. |
