Software is now operated by models as well as people. Models can use a computer the way people do, by looking at a screen, moving a cursor and typing, and browser agents finish whole tasks on websites without a person clicking through them.
That progress created a choice every team building agents now faces. A web agent works inside a browser and reads the page's structure; a computer use agent works across a whole desktop through screenshots, mouse and keyboard. Use a web agent when the job lives on websites, and a computer use agent when it touches desktop apps, local files or several programs at once. This guide covers how each one works, which jobs fit each, and what both need to run in production.
What Is the Difference Between Computer Use Agents and Web Agents?
Both agents run the same loop: look at the current state, decide the next action, take it, and check the result. The difference is the surface the agent can see and act on.
A web agent reads a page through its DOM, the live tree of elements a browser builds from HTML, and through the accessibility tree, which names each control, its role and its state. It clicks a button by reference instead of guessing pixels, and falls back to a screenshot only when the page structure does not describe what it needs. Anthropic's browser use tool works this way: Claude reads the page's structure and acts on element references.
A computer use agent sees the screen as an image. Anthropic's computer use tool gives Claude screenshot, mouse and keyboard control of a desktop, so it can operate any program a person could, from a spreadsheet to an accounting app. The trade is precision: small text, overlapping windows and moving interfaces make each step harder to judge from pixels.
Why Does the Choice Matter Now?
Desktop automation used to mean robotic process automation: scripts that replay recorded clicks and break when a screen changes. Browser automation meant hand-written selectors in frameworks such as Playwright. Both followed a fixed path, so every layout change sent an engineer back to fix the script.
Models changed that. An agent now decides each step from what it sees, so the same agent can handle a page it has never visited. The question moved from whether a job can be automated to which surface the agent should work on, and that choice sets cost, speed and reliability.
Anthropic points tasks that stay inside webpages to its browser tool and keeps the computer tool for work that needs a whole desktop. Its guidance is to read the page structure before taking a screenshot, because a tree read of a typical page often costs fewer input tokens and gives the model references it can act on right away.
Isolation matters for both. Anthropic recommends running either tool in a dedicated virtual machine or container with minimal privileges and internet access limited to an allowlist of domains, because every page an agent reads is untrusted input.
When Should You Use a Web Agent?
Use a web agent when every step of the job happens on a website. These are the jobs teams hand to web agents first:
Portal downloads. The agent signs in to supplier and billing portals and downloads last month's invoices. Finance gets every file in one folder without anyone logging in to each site.
Price and availability checks. It walks booking and checkout flows, picks dates or options, and records the current price. You see the live number, not a cached one from last week.
Multi-step forms. It fills quote requests and applications across many sites, reading each field by its label. A form that moves a field to a new spot does not break the run.
Research across sites. It reads company pages, docs and news, then returns a structured summary with the source page for every line. A person checks the summary instead of the twenty tabs behind it.
Web app testing. It clicks through your own product the way a customer would and reports what failed. Releases get checked on real pages, not only in unit tests.
When Do You Need a Computer Use Agent?
Use a computer use agent when any step leaves the browser. These are the jobs that need a full desktop:
Desktop software with no API. The agent enters records into an ERP or accounting app that only has a graphical interface. Work that used to need a person at a keyboard runs on a schedule.
Work across several programs. It opens a downloaded file, updates a spreadsheet and moves the result into a report. One agent handles the whole chain instead of three scripts glued together.
Files on disk. It saves, renames and uploads files through a real file system and terminal. Anything a person does with a folder, the agent can do too.
Canvas and remote screens. It works on interfaces the page structure does not describe, such as canvas editors and embedded remote desktops. Pixels are the only view available, so an agent that reads the screen is the right tool.
Many real jobs mix the two. The pattern that holds up is a web agent for the browser steps and a computer use agent only for the step that needs the desktop.
What Role Does Solari Play?
Solari is the infrastructure both kinds of agent run on: cloud browsers, sandboxes and desktops from one API. Your agent brings the model and the logic; Solari gives it a place to work.
Browsers for web agents at fleet size. Solari browsers are ready in 8 ms, and with Solari you can launch up to 100,000 browser sessions at once. Your web agent checks every portal or price on a list in one pass, connected through Playwright or any client that speaks the Chrome DevTools Protocol.
Logins that stay signed in. A Solari profile saves a session's login, cookies and site data, so your agent signs in once and every later run starts signed in. Profiles use Playwright's storageState format, so a login you already have uploads as is.
Sites that block automation. Turn on stealth for sites with bot defenses and route that traffic through a proxy. Every other site runs on the default browser, which is the fastest to start.
Full desktops for computer use agents. A Solari desktop is a full Linux computer with a graphical screen that starts usually in about a second. Your agent gets mouse and keyboard input, screenshots, a shell, files and a clipboard, with Google Chrome installed for the steps that need a browser.
A person who can take over. Every Solari desktop comes with a live view you can watch and control, so a person can step in mid-task. Record a browser session and replay it later to see what your web agent did; recording captures input values by default, so leave it off for sessions that type secrets.
Pause between steps. Pause a desktop and its state is saved, so the next connect picks up where the agent stopped. A paused desktop does not count against your limit on running VMs, so an agent waiting on a person holds no slot.
Both kinds of agent also write code to parse files and check results. Solari sandboxes run that code on isolated machines with no screen, with browsers, sandboxes and desktops on one account and one bill.
Where Are Computer Use and Web Agents Going Next?
Today most teams pick one kind of agent per project. Over the next year that changes in four ways:
One agent, both surfaces. Model vendors already ship browser and computer tools side by side. Agents will pick the surface per step, reading structure on websites and switching to pixels only when a step leaves the browser.
Structure first, pixels as fallback. Screenshot-only agents will keep improving. Web agents that read the page will still be cheaper per step for browser work, so most fleets will run both.
Long tasks on paused machines. Agents that wait on tool calls or people will pause their desktops between steps. A task that runs for hours will hold a machine only while it works.
Isolation by default. Every agent task will get its own browser or VM with a domain allowlist. Teams will treat what an agent reads on the open web as untrusted from the first run.
What stays human is judgment: which jobs to hand over, approval of actions with real consequences such as payments, messages and accepted terms, and ownership of the result. The teams that build that approval step in from the start will spend their people on decisions and hand the clicking to agents.
Get Started With Solari
Start with one web job, such as downloading last month's invoices from your vendor portals, and add a desktop only for the step that leaves the browser. Solari is the fastest and cheapest way to run cloud browsers and VMs for computer use and web agents. Get started with a free account and start running your agents today.
