Computer use and agents that control browsers: real cases in 2026
Agents that see the screen and move the mouse went from demo to production. Where they work, where they don't.
April 1, 2026 · Lixto Labs Team · 1 min read
What was sci-fi is here
In late 2024 Anthropic launched Computer Use and OpenAI launched Operator. By 2026 these models are significantly more reliable, and Mexican companies use them to automate tasks that previously only a human at a screen could do.
What they do
Give the agent access to a virtual machine and an objective. It takes screenshots, identifies UI elements, moves the mouse, fills forms, and navigates apps — like a person.
Where it works
- Bulk uploads in legacy systems: old ERPs without APIs. We went from 6-hour manual processes to 20 minutes unsupervised.
- Competitor and price research: navigate competitors' sites, capture prices, download catalogs, build reports overnight.
- Supplier onboarding: from an email, register the supplier in SAP/Odoo, upload docs, mark approval flow tasks.
Where it doesn't (yet)
- Sites with strict CAPTCHAs or bot detection: many sites detect agents and block them.
- 2FA-required tasks: solvable but complex and risky.
- Critical processes without supervision: still 5-15% error rate. If you need 99.9%, keep human-in-the-loop.
Practical advice
Computer Use shines in low-risk, high-volume automations on your own internal systems (your ERP, CRM, internal tools). When there's no API and a custom integration is expensive, computer use is the biggest productivity unlock we've seen in years.