The next agent surface is the computer itself. OpenAI is adding computer environments to the Responses API and Agents SDK. Google DeepMind is moving Project Mariner toward the Gemini API. The direction is clear: models are being trained and wrapped to operate tools, files, browsers, and command lines.
That can help a business. It can also create expensive mistakes if the work loop is loose.
What changed
| Source | Capability | Practical meaning |
|---|---|---|
| OpenAI Responses API | shell tool plus hosted container workspace | Agents can inspect files, run commands, and produce artifacts in an isolated workspace |
| OpenAI Agents SDK | sandbox execution and file/tool harness | Developers can give agents controlled environments instead of raw machines |
| Google Project Mariner | browser agents on virtual machines | Agents can research, plan, enter data, and repeat browser workflows |
Good first use cases
Computer-use agents are best when the screen work is repetitive and the risk is bounded.
- collect public information from a set of pages
- compare data across vendor portals
- fill a draft form without submitting it
- reconcile browser-visible records against a spreadsheet
- prepare a report from files in a controlled workspace
- test a website workflow and return screenshots
They are a bad first step for payroll, banking, public posting, or anything that submits irreversible changes.
The minimum production wrapper
| Control | Why it exists |
|---|---|
| isolated workspace | the agent should not see the whole machine |
| allowlisted domains | browser work should stay inside known surfaces |
| action log | every click, command, and file write needs a receipt |
| budget cap | long loops can spend money fast |
| stop condition | the agent needs a clear definition of done |
| human approval | sensitive actions need explicit consent |
| replay artifact | screenshots or files prove what happened |
What this means for small businesses
The first wave of value will not be "let the agent run the company." It will be small browser tasks that a person hates doing and can easily review.
That is enough. A weekly two-hour admin loop becomes a 10-minute review. A lead researcher turns scattered pages into a source-linked brief. A website QA pass returns screenshots and exact repro steps.
The agent should save attention, not hide the work.
