| name | browser-agent |
| description | Build, run, and review browser-using AI agents with Playwright, Browser Use, MCP browser tools, or hosted browser runtimes. Use for web navigation agents, form-filling, UI task automation, screenshots, DOM/vision checks, auth/session handling, and browser-task reliability reviews.
|
| version | 1.0.0 |
| author | Zijian Guo |
| license | MIT |
| platforms | ["linux","macos"] |
| prerequisites | {"commands":["python3"]} |
| metadata | {"hermes":{"tags":["agent","browser","playwright","browser-use","web-automation","ui"]}} |
browser-agent - Browser automation agents
Use this skill when an agent needs to operate a real website rather than only
fetch static HTML.
Research Signal
Browser agents are a fast-growing public category. Browser Use reports real-world
browser task benchmarks and exposes custom tools, hosted/cloud execution, MCP,
and authentication patterns. The production gap is reliability: screenshots,
DOM state, session handling, CAPTCHAs, consent screens, and side effects must be
handled deliberately.
Workflow
- Classify the task: read-only browse, authenticated browse, form fill, purchase
or destructive action.
- Use Playwright for deterministic UI tests; use browser-agent frameworks when
high-level planning across pages is needed.
- Capture screenshots, current URL, visible text, console errors, and network
failures at each checkpoint.
- Require explicit approval before submitting forms, purchasing, sending
messages, or changing account data.
- Convert repeated browser failures into eval cases.
Script
~/.hermes/skills/browser-agent/scripts/browser_env_check.py
Readiness Checks
| Check | Why it matters |
|---|
| Python packages | Detects Browser Use and Playwright availability |
| Node commands | Detects Playwright CLI/npm availability |
| Browser binaries | Indicates whether a local browser can run |
| Environment variables | Warns about API keys needed by hosted runtimes |
Pitfalls
- Static HTTP fetch is not a substitute for JS-heavy pages.
- CAPTCHA and bot-detection handling may require hosted browser infrastructure.
- Never let a browser agent auto-submit high-impact actions without raw-action
approval.