Skip to main content
All projects

Computer-use agent · AI Agents

Computer Using Agent

Legacy bank back-office screens, admin consoles and vendor portals rarely ship an API. Point cua at one with a goal in plain English and Gemini computer use works the task out on the real UI, once. What it learned is saved as a typed, reviewable, versioned capability file, and every run after that is plain code driving verified locators — fast, cheap and identical every time. When a replay meets something it has not seen, a person takes over the same live browser (or the AI decides, with --ai) and the capability remembers the decision as a new draft version, so code handles it alone next time. The split is deliberate: the model decides what a screen means, code holds the guarantees it cannot be trusted with.

1 / 1

Highlights

  • Record once with Gemini computer use, replay forever from versioned YAML — the replay path never calls a model
  • Anything unexpected escalates to a human in the same live browser, and the decision is written back as a new draft version
  • Code enforces what the model cannot: domain allowlist, approval before risky steps, and a risky step is never retried or repeated after a restart
  • Secrets marked {name:secret} are read from the environment and never reach the model, the artifacts, the logs or the screenshots
  • Site-agnostic — iframes, shadow DOM, pop-ups, new tabs, row-scoped locators and cross-domain SSO; CAPTCHAs are left to a human by design
  • A ~3-minute pytest suite drives the real recorder and replayer against a local legacy site with scripted stand-ins for every AI part — no test calls the real model

Tech stack

  • Python
  • Playwright
  • Gemini
  • YAML
  • uv
  • pytest