The auditor's question sounded easy: on the morning the vendor records changed, who changed them? The change log named a service account, and the service account belonged to the agent that copies orders from a spreadsheet into the CRM. No person on the team could say what that agent had seen on screen before it clicked Save.
That is the gap in most computer-use setups: a log of API calls shows what the agent sent, not what it saw. Solari closes it with two habits in its Computer Desktops: every job gets a desktop per run, an isolated Linux machine of its own, and that desktop can be recorded from the moment it starts, with a live view for a person to watch or step in. Here is how to set that up on Solari, and the rule about secrets that matters most.
Why Each Run Gets Its Own Desktop
An audit trail is only as clean as the machine it describes, so the first decision is where the agent runs. On Solari, each create call returns a fresh desktop with a stable id, usually ready in about a second. Starting a new desktop per job costs almost nothing in time, so there is no reason to share a machine between runs.
Isolation sits underneath that. Every Solari session runs in a Cloud Hypervisor microVM with a dedicated kernel, the kind of boundary a hypervisor draws between guests. A run cannot inherit files, cookies or open windows from the run before it.
For an auditor, that becomes a simple mapping: a VM id stands for a single job, and nothing else ever happened on that machine. With the machine settled, the next step is capturing what happens on it.
How to Record a Desktop Session
Recording is a flag set when the desktop is created, so it belongs in the create call rather than in a later step. The flag is record: true on Solari desktops, and since the September 6 release a desktop created that way returns its recordingUrl on create, with no second lookup.
The sketch below starts a recorded desktop, stores the playback link with the job, runs the agent and destroys the machine at the end.
There is an edge case worth knowing: an in-progress recording ends at a revert(), because a revert replaces the machine, and a recording started after it uploads to the same playback URL. Once the video exists, the work is making it useful as evidence.
Turning a Recording Into an Audit Trail
A recording on its own is footage; an audit trail is a chronological record that lets someone rebuild the sequence of events. Solari desktops are built for both halves of that job: a run can be watched live, taken over when a human is needed, and opened as a replay afterwards for review.
The strongest setup pairs the video with the agent's own action log. Each entry should answer when, where, who and what, the four questions at the heart of application logging, so every model step carries a timestamp, the VM id, the job id and the action taken. A reviewer can then jump from a log line to the moment in the recording where the agent clicked Save.
Many computer-use jobs also open a cloud browser for the web steps. Those sessions record with their own flag, and a browser replay downloads as a line-by-line log that a team can keep under its own log management policy. Evidence this detailed brings its own risk, which is the next thing to plan for.
What to Keep Off the Screen
A recording shows whatever the screen showed, so anything typed or displayed becomes part of the record. Solari recording captures input values by default, so replays that touch passwords or payment data need a short list of readers and a purge schedule.
The practical rule is to keep secrets out of the recorded desktop. Sign-ins belong in a step that is not recorded, and the agent should work with accounts scoped to the job. The live view needs the same care: its streamUrl is a credential that lets the holder drive the VM as root, so it belongs in a secrets store, never in a ticket.
With the risks contained, the last question is what a recorded fleet costs to run.
What Recorded Desktops Cost
Desktops bill by the time they run, and plan limits decide how many can run at once. The figure below shows concurrent desktops per plan.
On the Professional plan, a small desktop with its live screen runs at $0.060 per desktop hour, and the live screen on its own adds $0.02 per hour. A paused VM does not count against the plan's limit on running VMs, so a recorded job waiting on a human review holds no slot while it waits.
Choosing the Right Setup
A computer-use agent is auditable when each run has its own machine, its own recording and a log that points into that recording. Create a fresh desktop per job, set the recording flag at create, store the playback link with the job record, and keep secrets and stream links off the screen. The companion guide on what a computer-use desktop environment is covers how agents drive the screen, and the Solari introduction covers the browsers and sandboxes that sit beside it.
Get Started With Solari
The first job to try is a short recorded run: start a desktop with the recording flag, let the agent finish a single task, then watch the replay next to its action log. Solari is the fastest and cheapest place to run it, with the desktop up in about a second as above and billed at the Professional rate above, from the same account as its browsers and sandboxes. Create a free Solari account and the Free plan's credits cover the first recorded desktop.
