How to hand off computer tasks to an AI agent

By the end of this guide you will have handed one real, repetitive computer task to an AI agent, watched it complete the task, and set limits so it cannot do more than you intended. An AI agent that uses your computer takes a plain-language instruction and then clicks, types, browses and produces files for you, instead of only replying with text.

The choice that matters most is where the agent runs: inside your own browser, in a hosted cloud workspace, or on your machine with system-level control. This guide compares six agents from our Automation and agents category and walks through a safe first handoff.

Key takeaways

• Start with one low-risk, repeatable task, describe the outcome in plain language, watch the agent do it once, and only then let it run unattended.
• The main decision is where the agent works: your own logged-in browser, a hosted cloud agent, a local framework with system access, or a phone.
• Agents that use your logins or control your computer can act as you. Limit what they can reach, and keep a human approval step for payments, sending messages and deleting things.
• Usage is metered in credits or API steps on most plans, so check how the cost grows with task complexity.

What you need before you start

• One well-chosen task. Something repetitive and low-stakes with a clear finish, such as collecting prices, filing a report or updating a list.
• Written steps. A few lines describing what "done" looks like, so you can judge the result.
• The right accounts. Agents that work in your browser use the sites you are already signed in to. A hosted agent needs its own account.
• A safe test setup. A test account, a copy of a file or a low-value record to try first.
• Time. A first handoff takes under an hour to set up and watch.
• Difficulty. Easy for a browser extension or hosted agent, harder for the self-hosted framework, which needs Python and your own model key.

Which kind of AI tool fits this job

Pick by where you want the agent to act and how much control you want to keep.

An agent inside the browser you already use. Composite runs in your existing browser: you press a shortcut on a website, describe the task, and it clicks and types using your logged-in sessions. It can run in background tabs and reach any site you can sign in to. The listing names Chrome, Edge, Brave and Arc as supported desktop browsers.

A hosted, general-purpose agent. Manus does the work behind a request: research, slide decks, websites, images and scheduled automations, available on the web, in mobile and desktop apps, through Slack and by API. Magica plans the steps for an outcome and uses a browser, a cloud computer, files and connected apps. Cortex Workspace runs long tasks in a desktop app and on the web, automating browser tasks such as filling forms and downloading files, and creating Word, Excel and PowerPoint files.

A local framework with full control. GenericAgent is an open-source framework that lets a model operate a local computer through code execution, file access, browser control, keyboard and mouse. It is free and MIT-licensed, but you supply your own model key and it gives the model system-level control, so it needs care.

An agent for a phone. AGI, Inc. builds agents that see the screen and tap and type to finish tasks in any Android app. Android is available now, iOS is in early preview, and the company also sells a hosted API for browser and OS control agents.

Step 1: Pick a task that is safe to hand over

Choose the task that proves the idea with the least risk.

1. List the repetitive computer tasks you do each week.
2. Cross out anything that moves money, sends messages to customers, deletes data or changes permissions.
3. Pick one task with clear start and end points, such as gathering information from several pages into a document.
4. Estimate how long it takes you now, so you can compare later.

Check it worked by being able to describe the finished result in two sentences.

Step 2: Choose where the agent will run

Match the environment to the task and to how much access you are willing to give.

1. If the task lives in websites you are logged in to, consider a browser agent like Composite.
2. If you want a finished deliverable such as a report, slides or a spreadsheet, consider a hosted agent such as Manus, Magica or Cortex Workspace.
3. If you want local, scriptable control and are comfortable with Python, consider GenericAgent.
4. If the task lives in an Android app, consider AGI, Inc.
5. Check the platform. Composite is a Chrome extension, Cortex Workspace lists web and Windows, and Magica lists web, iOS, Android and an API.

Check it worked by confirming the tool runs on the device and browser you will use.

Step 3: Set up the account and the access

Give the agent only what it needs.

1. Create the account or install the tool. GenericAgent needs a clone of the repository, your own model API key, and Python 3.11 or 3.12.
2. Use a test or low-privilege account where you can. Composite acts through your own signed-in sessions, so anything you can do on a site, it can try to do.
3. Connect only the apps the task needs. Magica lists connections to Google, GitHub, Slack and Notion with permission controls.
4. Check the plan. Manus has a Free plan with daily credits, one concurrent task and a limit on scheduled automations, and Composite has a Free plan with a lower request allowance than Pro.

Check it worked by running a trivial test, such as asking the agent to open a page and read the title.

Step 4: Write the instruction clearly

Vague goals lead to wandering agents, so say what you want and where to stop.

1. State the outcome: "collect the prices of these products from these sites and put them in one table".
2. Name the sites, files or apps to use, and the ones to avoid.
3. Say what the agent must not do, such as submit forms, purchase anything or send email.
4. Tell it to stop and ask when it is unsure. Cortex Workspace says its browser agents can pause to ask for input at a decision they cannot make, and GenericAgent can ask the user for confirmation.

Check it worked by reading the instruction as a stranger would and confirming nothing important is ambiguous.

Step 5: Run it once while you watch

The first run is for observation, not for saving time.

1. Start the task and keep the window visible.
2. Watch each step and stop the agent if it goes to the wrong place or tries something you ruled out.
3. Note where it hesitated or made a wrong choice.
4. Keep a record of what it changed, such as a list of edited rows or downloaded files.

Check it worked by comparing the result with what you would have produced yourself.

Step 6: Review the result and fix the instruction

Treat the output as a draft from a new assistant.

1. Check every figure, name and link that matters against the source.
2. Look for steps the agent skipped or invented.
3. Rewrite your instruction to prevent each error you saw, then run it again on a fresh copy.
4. Compare the time and credits used with doing the task by hand.

Check it worked by getting a correct result twice in a row on similar input.

Step 7: Add supervision, then schedule or expand

Only after two good runs should you reduce supervision.

1. Keep human approval for anything with real consequences: payments, outgoing messages, deleting data.
2. Schedule recurring runs if the tool supports it. Manus lists scheduled automations, and GenericAgent lists scripted scheduled web summaries as a use.
3. Set a usage limit or watch the credit dashboard. Cortex Workspace has a credit usage dashboard.
4. Review a sample of the output regularly, and stop the automation if the sites or apps change.

Check it worked by reviewing a week of runs and finding no unexplained changes.

Tools for this job compared

Tool	Good fit for	Pricing	Stated limits
Manus	Hosted agent for research and deliverables, on web, desktop, Slack and API	Freemium, paid from $20/mo (checked 24 Oct 2026)	Free plan has 300 daily credits, 1 concurrent task and 2 scheduled automations; some features in beta
Composite	Repetitive tasks in your own signed-in browser	Freemium, paid from $20/mo (checked 24 Oct 2026)	Desktop browsers only; Free plan has a lower monthly AI request allowance than Pro; you must be signed in
Cortex Workspace	Browser automation and Office files on web and Windows	Freemium, paid from $4.99/mo (checked 24 Oct 2026)	Heavier models and longer steps use more credits; premium model usage capped on some plans
Magica	Multi-step outcomes with browser, cloud computer and apps	Paid, from $14.99/mo (checked 24 Oct 2026)	Credits removed when a subscription ends; payments non-refundable; Teams needs at least 5 seats
GenericAgent	Local, scriptable computer control for technical users	Open source	Gives the model system-level control; needs your own model key and Python 3.11 or 3.12
AGI, Inc.	Android app tasks and developers building agents	Free trial, paid from $19/mo (checked 24 Oct 2026)	iOS is an early preview; MCP sessions capped at 200 steps; benchmark figures are company-reported

Prices and plan limits change, so confirm the current terms on each vendor's site before you subscribe.

Common mistakes to avoid

• Starting with a high-stakes task. An agent that mis-clicks on a payment or a customer email causes real damage. Fix it by starting with read-only or reversible work and keeping approval steps.
• Giving the agent broad access. An agent working through your logged-in sessions can do anything you can. Fix it by using test or low-privilege accounts and connecting only the apps the task requires.
• Trusting the result without checking. Agents can skip steps or misread pages. Fix it by spot-checking figures against sources, especially for finance, legal or customer-facing work.
• Ignoring credit costs. Credits and steps are consumed according to task complexity, so loops and long tasks add up. Fix it by testing small and watching the usage dashboard.
• Running a local agent without safeguards. GenericAgent's own listing warns that it gives the model system-level control of your computer. Fix it by running it on a machine or profile with nothing you cannot afford to lose.

How we put this together

This is an evidence-based guide built from each tool's listing in our directory, which is compiled from the vendor's own site and pricing page. It is not a review based on using the tools. We don't rank tools, and nobody paid to be included. Some outbound links may be affiliate links; see our disclosure.

Frequently asked questions

Is there a free AI agent that can use my computer?

Yes, with limits. Manus and Composite have Free plans, Cortex Workspace lists a Free tier, and GenericAgent is free and open source, though you pay for your own model API use. Free plans cap daily credits, requests or concurrent tasks.

Is it safe to let an AI agent control my computer?

Only with precautions. Browser agents use your logged-in sessions, and local frameworks can control the whole system. Use test accounts, limit what the agent can reach, require approval for risky actions and review what it did.

Which agent works on a phone?

AGI, Inc. lists Android as available, with iOS in early preview. Magica lists iOS and Android apps and Manus lists mobile apps, but the listings do not say they operate other apps on the phone. Check each vendor's site for current capabilities.

Can an AI agent do a task while I do something else?

Some can. Composite runs in background tabs, Manus runs scheduled automations and Cortex Workspace runs browser agents in parallel. Keep supervision until you trust the results.

How accurate are computer-using agents?

Not guaranteed. AGI, Inc. itself says its MCP release is early and that the agent will sometimes miss. Treat results as drafts and verify anything that matters.

Compare more options in our Automation and agents category.