Solari for Computer-Use Agents: Desktops and Recorded Audit Trails

Computer use has moved from research demo to product. OpenAI, Anthropic and Google DeepMind have all built agents that use a computer through its screen, reading pixels and answering with clicks and keystrokes the way a person does.

That shift creates two needs at once: a screen the agent can operate, and a record of what was on that screen when it acted. Text logs show what the model intended; only a recording shows what the machine displayed. This guide covers why computer-use agents need their own desktops, what you can do with a recorded one, how Solari runs them, and where these agents go next.

Why Does a Computer-Use Agent Need Its Own Desktop?

A computer-use agent needs a real graphical session, because it works from screenshots and answers with mouse and keyboard actions. Anthropic's computer-use docs recommend running the model in a dedicated virtual machine or container with minimal privileges, so a wrong click stays inside one machine.

The record changes too. An agent that called APIs left a text trail of requests and responses. An agent that acts on a screen leaves visual evidence, and an audit trail, in NIST's definition, is a chronological record that reconstructs the sequence of events behind an operation. For a computer-use agent, the screen is the part a text log misses.

Regulation points the same way. The EU AI Act requires high-risk AI systems to allow the automatic recording of events over their lifetime, so a team that ships agents into regulated work needs evidence it can replay, not only a transcript.

What Can You Do With a Recorded Desktop?

A desktop per run plus a recording of its screen turns a computer-use agent into something a reviewer can check. These are the jobs teams set up first:

• Give every run a clean machine. Each task starts on a fresh desktop and ends with destroy, so a file or a cookie left by one run never appears on the next agent's screen. The recording then shows a single agent doing a single task, which is what makes it evidence.

• Record the screen from the first frame. Create the desktop with recording on and store the playback link beside the run's model calls. A reviewer reads what the model meant and watches what it did, side by side.

• Review a workflow before it runs unattended. Watch the replay of a new workflow end to end before you let it run on its own. Once it holds up, review a sample of runs each week instead of every run.

• Rewind and retry without losing the record. Take a snapshot before a risky step and revert if the step fails. Each attempt gets its own recording, so the footage shows every try, not only the one that worked.

• Record the browser steps too. When part of the job runs in a cloud browser, that session records as well. Its replay downloads as a line-by-line log of everything that happened on the page, which you can search when a question comes up months later.

• Watch live and take over. Open the live view to follow a run as it happens, and take the mouse when the agent gets stuck. The view works like VNC desktop sharing, in a browser tab.

Each of these depends on the desktop, the recording and the live view behaving the same way on the thousandth run as on the first. That is the part Solari runs for you.

What Role Does Solari Play?

Solari runs browsers, sandboxes and desktops from one API and one account. The agent brings the model and the logic; Solari gives it a screen, a recording and a clean machine for every run.

• A desktop in about a second. A Solari desktop is a full Linux computer with a graphical screen, ready usually in about a second, that your agent drives from the SDK with mouse, keyboard, screenshots and a shell. A fresh machine per task adds almost nothing to the run.

• A kernel of its own. Each Solari session runs on its own Cloud Hypervisor microVM with a dedicated kernel, so runs stay apart at the machine level. What one agent leaves behind cannot reach the next.

• Recordings with no second lookup. Create a desktop with record: true and Solari returns a recordingUrl on create and every time you fetch the desktop. A recording ends at a revert, since a revert replaces the machine, so start a new one after each revert; it uploads to the same playback URL.

• Browser replays on the same account. Session recording on a Solari cloud browser is one flag at launch. You open the replay in the console or download it as NDJSON to keep with the run.

• Pause instead of paying for idle. A Solari desktop pauses when its idle timeout passes, and if you set none, the timeout "defaults to 30 minutes". A paused desktop does not count against your limit on running VMs; same-host resume takes about 1.3 seconds on staging, five trials, and parking freezes the CPUs for short waits on a model response while still counting against that limit.

Starting a recorded desktop takes a few lines:

import { DesktopClient } from "@solarisdk/desktop"

const desktops = new DesktopClient({ apiKey: process.env.SOLARI_API_KEY! })
const desktop = await desktops.create({ template: "default", record: true })
await desktop.connect()
await desktop.health()
const png = await desktop.screenshot({ format: "png" })

import { DesktopClient } from "@solarisdk/desktop"

const desktops = new DesktopClient({ apiKey: process.env.SOLARI_API_KEY! })
const desktop = await desktops.create({ template: "default", record: true })
await desktop.connect()
await desktop.health()
const png = await desktop.screenshot({ format: "png" })

import { DesktopClient } from "@solarisdk/desktop"

const desktops = new DesktopClient({ apiKey: process.env.SOLARI_API_KEY! })
const desktop = await desktops.create({ template: "default", record: true })
await desktop.connect()
await desktop.health()
const png = await desktop.screenshot({ format: "png" })

Solari keeps replays from one day up to 90 days, so match retention to how often you review runs.

A Solari desktop costs $0.060 per hour at Solari's best self-serve rate, and the live screen adds $0.02 per hour. Solari sandboxes share the same lifecycle without a screen, for the code your agent writes and runs, and Solari is almost 3x cheaper than E2B per sandbox hour.

How Do You Keep Secrets Out of the Recording?

A recording that shows everything also shows anything secret, so treat it as sensitive data from the start. Solari browser recording captures input values by default, and a desktop recording is the screen itself: a customer record on a form or a key pasted into a terminal ends up in the footage.

• Mask before you store. The OWASP logging guidance applies to recordings as to any log: passwords, tokens and payment data are removed or masked before anything is kept. Keep agents away from those fields where you can, so there is less to mask.

• Restrict and purge. Limit who can open replays of sessions that touch passwords or payment data. Purge them on the schedule your compliance team sets.

• Treat the live view as a key. A desktop's streamUrl is a credential: anyone holding it can watch and drive the VM as root until it is destroyed or expires. Keep it out of chats, tickets and logs.

Where Are Computer-Use Agents Going Next?

Today most computer-use agents run one task on one screen with a person nearby. Over the next year that changes in four ways:

• From one screen to fleets. Back-office jobs that live in apps with no API move to agents running many desktops in parallel, each on its own machine.

• Recordings as standard evidence. Reviewers and auditors will ask for the footage of an agent's run the way they ask for logs today.

• Agents that review agents. A second agent will watch replays and flag the steps worth a person's attention, so review scales with the fleet.

• Longer runs. Agents will work for hours across pauses and resumes, and the recording will follow the task rather than a single session.

What stays human is approval: which actions an agent may take on its own, what counts as a consequential step, and the decision when the footage shows something unexpected. The teams that do this well spend their people on those calls and hand the clicking to agents.

Get Started With Solari

Start with one recorded desktop running a task your agent already handles, and watch the replay end to end before the fleet grows. Solari is the fastest and cheapest way to run cloud browsers and VMs for computer-use agents. Get started with a free account and start running your agents today.