AgentComputer

Computer

Computer gives your agent a computer of its own for each session. On it, the agent works the way a person at a desk would: it opens files, runs calculations, writes and runs its own code, and makes new files. So instead of only describing the work, it can do it and hand over the result.

How it works

One computer per session. Each conversation, form run or workflow run gets its own computer, separate from every other session. It starts the first time the agent needs it and stays with that session: come back tomorrow, and the files are still there.

The agent decides when to use it. Like any tool, the agent reaches for its computer when the task calls for it: a spreadsheet to analyse, a report to make, a calculation to get exactly right. Many replies need no file at all, and the agent answers in the chat as usual.

It works with three kinds of files:

  • Your knowledge. The files you gave the agent in Knowledge, to read and build on. The agent never changes them.
  • The session. What people shared in this conversation or run, and what was said and made in earlier turns or steps.
  • Its own work. Scripts, drafts and anything in between, which stay private, and the finished files it hands over.

Finished files arrive with the reply. When the agent hands over a file, it’s attached to its answer, in the app, on your published chatbot and in a form’s results. People preview it in the browser and download it. When the agent changes a file later, say “make it shorter and add a chart”, the change becomes a new version of the same file, and earlier versions are kept.

In a workflow, the steps share one computer. A later step can open what an earlier step made, so one step can clean up a spreadsheet, the next can analyse it, and a third can turn the findings into a deck.

What it can do

The computer runs Python and Node.js, with the libraries for documents, data and charts already installed, plus tools for converting and rendering files. In practice, that means:

It can…For example
Read the files people shareSpreadsheets, CSV files, PDFs, Word documents, PowerPoint decks, images
Work through dataClean up and combine lists, do the maths exactly, find trends, run statistics
Make documents people useWord documents, PDFs, Excel workbooks with live formulas, PowerPoint decks with speaker notes
Show things visuallyCharts, images, web pages and small interactive pages such as calculators or quizzes
Change a file’s formTurn a Word document into a PDF, a spreadsheet into a chart, a set of files into one document

It also looks at what it made before handing it over, the way a person glances at a page before sending it.

Who the session is with

What the computer can reach depends on who the agent is working with, so a public AI tool can never be used to take your material.

The session is with…The internetYour knowledge files on the computer
You and your team, in the appYesYes
A workflow on a schedule or through the APINoYes
Visitors on a published chatbot or formNoNo: the agent still answers from your knowledge, as always

The computer never holds passwords or keys.

Put it to work

Switch on Computer in the agent’s Tools. New agents start with it on.

  • Name the finished result in the instructions. “At the end of the session, give the client their plan as a one-page PDF” works better than “help the client plan”. Say when it should not make a file, too: “Answer short questions in the chat.”
  • Point to files in plain words. “Use the proposal template in your knowledge” or “start from the spreadsheet the client shares”. The agent finds them.
  • Shape the session around the result. A few patterns builders use:
    • A file in, a better file back: a client shares their numbers and gets a report with charts.
    • Answers in, a personal file back: a form or a conversation ends with each person’s own plan, report or workbook.
    • Steps that build one result: a workflow researches, analyses and then assembles the deck.

In a workflow, give a step an agent that has Computer and name the file in the step’s prompt. A step with no agent has no tools.

Limits

  • One command can run for up to 2 minutes in a chat and 5 minutes in a workflow.
  • One session can run commands for 30 minutes in all.
  • Each file the agent makes can be up to 50 MB, and the files it makes in one session up to 500 MB together.
👋
MindPal is a platform that helps you build AI agents & multi-agent workflows to automate business processes. Get started here!