wallacedobbs428/agent-evaluation

agent-browser - Browser Automation for AI Agents

When to use this skill

• Open websites and automate UI actions
• Fill forms, click controls, and verify outcomes
• Capture screenshots/PDFs or extract content
• Run deterministic web checks with accessibility refs
• Execute parallel browser tasks via isolated sessions

Core workflow

Always use the deterministic ref loop:

1. agent-browser open <url>
2. agent-browser snapshot -i
3. interact with refs (@e1, @e2, ...)
4. agent-browser snapshot -i again after page/DOM changes

agent-browser open https://example.com/form
agent-browser wait --load networkidle
agent-browser snapshot -i
agent-browser fill @e1 "[email protected]"
agent-browser click @e2
agent-browser snapshot -i

Command patterns

Use && chaining when intermediate output is not needed.

# Good chaining: open -> wait -> snapshot
agent-browser open https://example.com && agent-browser wait --load networkidle && agent-browser snapshot -i

# Separate calls when output is needed first
agent-browser snapshot -i
# parse refs
agent-browser click @e2

High-value commands:

• Navigation: open, close
• Snapshot: snapshot -i, snapshot -i -C, snapshot -s "#selector"
• Interaction: click, fill, type, select, check, press
• Verification: diff snapshot, diff screenshot --baseline <file>
• Capture: screenshot, screenshot --annotate, pdf
• Wait: wait --load networkidle, wait <selector|@ref|ms>

Verification patterns

Use explicit evidence after actions.

# Baseline -> action -> verify structure
agent-browser snapshot -i
agent-browser click @e3
agent-browser diff snapshot

# Visual regression
agent-browser screenshot baseline.png
agent-browser click @e5
agent-browser diff screenshot --baseline baseline.png

Safety and reliability

• Refs are invalid after navigation or significant DOM updates; re-snapshot before next action.
• Prefer wait --load networkidle or selector/ref waits over fixed sleeps.
• For multi-step JS, use eval --stdin (or base64) to avoid shell escaping breakage.
• For concurrent tasks, isolate with --session <name>.
• Use output controls in long pages to reduce context flooding.
• Optional hardening in sensitive flows: domain allowlist and action policies.

Optional hardening examples:

# Wrap page content with boundaries to reduce prompt-injection risk
export AGENT_BROWSER_CONTENT_BOUNDARIES=1

# Limit output volume for long pages
export AGENT_BROWSER_MAX_OUTPUT=50000

# Restrict navigation and network to trusted domains
export AGENT_BROWSER_ALLOWED_DOMAINS="example.com,*.example.com"

# Restrict allowed action types
export AGENT_BROWSER_ACTION_POLICY=./policy.json

Example policy.json:

{"default":"deny","allow":["navigate","snapshot","click","fill","scroll","wait","get"],"deny":["eval","download","upload","network","state"]}

CLI-flag equivalent:

agent-browser --content-boundaries --max-output 50000 --allowed-domains "example.com,*.example.com" --action-policy ./policy.json open https://example.com

Troubleshooting

• command not found: install and run agent-browser install.
• Wrong element clicked: run snapshot -i again and use fresh refs.
• Dynamic SPA content missing: wait with --load networkidle or targeted wait selector.
• Session collisions: assign unique --session names and close each session.
• Large output pressure: narrow snapshots (-i, -c, -d, -s) and extract only needed text.

References

Deep-dive docs in this skill:

• commands
• snapshot-refs
• session-management
• authentication

Related resources:

• https://github.com/vercel-labs/agent-browser
• https://agent-browser.dev

Ready templates:

• ./templates/form-automation.sh
• ./templates/capture-workflow.sh

Metadata

• Version: 1.1.0
• Last updated: 2026-02-26
• Scope: deterministic browser automation for agent workflows