Chat or agent: what changes when AI runs commands itself?

Two ways of working: answers or manual steps

Most people know AI as a chat window: you type a question, the machine answers. For explanations, drafts and quick research, that is great. But everything that is supposed to come out of the answer, you then do yourself: the table in the right folder, the amended contract, the recalculated total. You copy, save, name and check.

An AI agent works differently. It is given a goal and a place and chooses the steps itself. In Biwak, this place is a working folder that you specify. The agent opens the PDFs, spreadsheets, Word files and earlier versions in it itself, and its results end up as files in the same folder. How agents work in general and which office tasks are suitable is covered in the article What is an AI agent?.

What a terminal command is, without jargon

Every Mac and every Windows computer has a text window in which you can write instructions to it: the Terminal, or PowerShell on Windows. Instead of clicking, you type a short command, the computer runs it and replies with text. Software developers have worked like this for decades, because a command is precise, repeatable and can be chained with others. Four examples from the log further down:

A chat can suggest such commands to you, but you would have to type them in yourself. An agent runs them, reads the computer's response and decides on the next step accordingly. That is the step from advice to action.

In software development, this way of working is already widespread. Anthropic describes Claude Code as a tool that reads code, edits files and runs commands; according to OpenAI, its Codex CLI runs “locally on your computer”. Even there, it is not yet everyday practice: in the Stack Overflow 2025 developer survey, 84% of respondents use or plan to use AI tools, but only 14.1% work with AI agents daily. Biwak transfers this way of working from source code to the office folder: instead of program files, it contains invoices, contracts and spreadsheets.

Same task, two routes: who does the manual steps?

Take an everyday task: a bank statement is available as a PDF, the transactions need to go into a table, and the total has to match the account balance. In ChatGPT, data analysis handles this. For many analyses, according to OpenAI, it writes Python code and runs it in an environment that works with the files made available to the session, i.e. with your uploaded copy. OpenAI states limits for uploads: 512 MB per file, around 50 MB for spreadsheets, up to 80 files in three hours, and three uploads a day on the free plan. You get the result as an answer in the chat, for example as a table or chart.

To be fair: ChatGPT can also recalculate the total if you ask. The difference lies not in how clever the model is, but in where the work happens. In the chat, the AI works with a copy in an environment at the provider. In the working folder, it works on the spot, next to all the other files it may still need: the price list, last month's spreadsheet, last week's contract.

A log, translated line by line

What does it look like when an agent works? Below are the commands of such a task. We ran them in the Terminal on a MacBook Pro on 24 September 2026, with the invented bank statement from our sample pack. On the left is what the computer saw, on the right what it means.

The interesting step is the fourth. The first check gave €14,894.37 instead of €16,679.37. It was not the table that was wrong, but the check: one payment reference itself contains a semicolon, exactly the character that separates the columns, and in row 8 the calculation picked the wrong column. We had not planned this error; it happened while we were putting the log together. It shows nicely what matters. A chat answer with a wrong total sounds just as convincing as one with the right total. An agent that recalculates in the folder notices the gap, because the number does not match the statement.

Why the folder matters more than the answer

In the chat, you decide what the model gets to see: whatever you paste in, upload or attach from a connected service such as Google Drive or OneDrive. Whatever you forget is missing, and the answer still sounds fluent. In the working folder, the agent looks for itself. It finds the price list next to the order and the template next to the draft. The folder is the context.

What this looks like with a larger volume is shown in the demo on our home page: a clause is replaced in 30 contracts, 28 are changed, two were already up to date, and a list of changes is created. The demo is a recreation without a model call; it shows the process, not a measurement.

Commands on your own computer: what protects you?

Anything allowed to run commands can also do damage. That is the price of real work, and we say so openly: the agent's commands reach as far as your user account on the computer. Biwak works in the folder you choose and saves the results there, but that is not a technical barrier for the rest of the computer. The terms and conditions expressly point this out in § 9. That is why there are these guard rails:

Your files stay in the folder on your computer. Biwak also stores the app's conversations on your computer, outside the working folder; they do not reach us, except in an error report or a message that you send yourself. The app does not report anything to us of its own accord: no telemetry, no usage statistics. For an answer, your question, the necessary context and the file contents used go via the Biwak relay server in Frankfurt am Main to the language model. Processing happens in the EU: on Microsoft Azure in the EU Data Zone, alternatively via Prem AI (contracting party PREM SA, Switzerland). Biwak does not store this content in the process and does not use it for training. What the model provider promises, and where this promise ends, is set out in the privacy policy, section 14 and on the security page.

What ChatGPT does better

• Questions, drafts, research: that is what a chat is built for, in the browser, on the desktop and on the phone. The Biwak desktop app practically does not search the web; web search is only available in the browser workspace.
• Teamwork: ChatGPT Business and Enterprise offer shared projects and administration for many seats. With Biwak, any number of colleagues work in the Rope Team plan with their own access from one quota; we do not promise a jointly edited workspace.
• OpenAI has agents too: Codex CLI works on the developer's own computer. In Business and Enterprise, ChatGPT Work can use approved files, programs and browser sessions on the computer; OpenAI itself stresses that the model still does not run on the device. According to OpenAI, the earlier “ChatGPT agent”, which worked on a virtual computer at OpenAI, no longer exists.
• Images: ChatGPT generates images, according to the pricing page also in Business and Enterprise. Biwak does not generate images and does not read scanned receipts without a text layer.

So the question is less “chat or Biwak?” than “what work needs doing?”. For questions, drafts and research, a chat is enough. As soon as a task consists of files that have to be read, changed, filed and checked, the working folder pays off. A comparison of the business ChatGPT plans with Biwak is available under Biwak or ChatGPT.

Try it yourself: a first task with a check

1. Download

   Biwak is available for macOS and Windows on the Download page. The app is free; model requests run via a Biwak account.

2. Create account

   After email confirmation, you get a one-off 100 trial credits, with no payment details. Create account.

3. Create a folder with copies

   For example three PDF invoices or a spreadsheet export, and choose this folder as the working folder in Biwak.

4. Write the task with a check

   Say how the result can be checked: “Transfer all invoices into a table and check whether the sum of the net amounts matches the individual invoices.”

5. Check the result

   The new file appears as a card with a preview. If something is wrong, reset the folder with “Undo”.

More ideas for getting started are in the article A good first task. After that there are two plans: Base Camp for €19 and Rope Team for €99 a month, both cancellable monthly (prices at a glance). Prefer to watch first? In a twenty-minute demo, we work on a real task from your day-to-day work.

How this comparison was made

This text was written by Biwak, i.e. a provider writing about its own product; it does not replace an independent test. The information on ChatGPT, Codex and Claude Code comes from the vendors' help and documentation pages and was checked against the wording on 24 September 2026. OpenAI's help pages cannot be retrieved automatically; for those, we read the Internet Archive copies of 2, 15, 17 and 18 September 2026. The information on Biwak comes from the privacy policy, the terms and conditions and the app's source code. The commands in the log ran on 24 September 2026 on a MacBook Pro with M1 Pro, with invented sample data and without a language model. The graphic of the manual steps is a diagram. We have not carried out a performance comparison between ChatGPT and Biwak. Please report errors to kontakt@biwak.ai.

Frequently asked questions