computer-use-agents-2026

Files (3)

• SKILL.md
• openai.yaml · agents/
• sources.md · references/

SKILL.md (first lines)

# Computer Use Agents (2026)

Use this skill when the agent must operate a real browser or desktop surface, not just call APIs or script a DOM.

## Workflow

1. Classify the surface first: deterministic browser automation, screenshot-based browser agent, or full desktop computer use.
2. Prefer deterministic browser tooling when DOM access is sufficient; escalate to computer-use only when visual grounding or arbitrary GUI control is required.
3. Select the runtime deliberately: OpenAI computer use, Anthropic computer-use plus bash/text editor, or a sandboxed browser-agent stack.
4. Run in an isolated VM, container, or controlled browser environment with minimal credentials, explicit checkpoints, and action logging.
5. Evaluate against representative UI tasks and benchmark-style scenarios before production rollout.

## Guardrails

- Separate test and production accounts, cookies, and browser profiles.
- Require human checkpoints for login, payments, destructive actions, and data export.
- Record screenshots, actions, and stop reasons for every validation run.
- Keep display size, action rate, and retry policy explicit so runs are reproducible.

## Output Requirements

- State whether the workflow uses browser automation, browser agenting, or full computer use.
- State the runtime and isolation model.
- State the approval/auth policy and one concrete stop condition.

From mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills-, MIT.

More from mdbabumiamssm/LLMs-Universal-Life-Science-and-Clinical-Skills-