Browser automation for AI agents is the practice of letting an agent operate a real browser while collecting enough evidence to verify what happened. The minimum useful loop is goal, page state, action, screenshot or DOM proof, recovery, and stop condition. Tangle Browser Agent packages that loop behind the bad CLI and SDK so teams can run natural-language browser tasks with evidence instead of hoping a model clicked the right button. Start with Browser Agent when the browser is the work surface.
The common failure is treating browser agents like a prettier wrapper around Playwright. Playwright is the control layer. The agent loop still has to decide what to do, inspect the result, recover from dead ends, and produce artifacts a reviewer can trust.
Safe Discovery And Install
npm install -g @tangle-network/browser-agent-driver
npx playwright install chromium
bad --help
bad run --help
bad run --goal "Verify the pricing page loads and capture evidence" --url https://tangle.toolsThe Browser Agent manifest exposes the package, CLI binary, safe help commands, and example run. It also points to related Tangle surfaces: Sandbox for machine runtime and Router for model selection.
DOM Mode, Vision Mode, Hybrid Mode
| Mode | Strength | Risk |
|---|---|---|
| DOM | precise selectors, text, attributes | misses visual-only state |
| Vision | sees layout, screenshots, rendered state | can misread hidden structure |
| Hybrid | combines DOM facts with screenshots | more expensive, but better for real apps |
Agents need both structure and evidence. DOM snapshots explain what the page exposes. Screenshots show what a user would see. Network and console logs catch failures that the DOM can hide.
How This Differs From Plain Playwright
Playwright is the browser automation substrate. Tangle Browser Agent adds a goal-driven loop, evidence capture, and CLI workflow for agent tasks. This puts it closer to agentic browser tools such as Browser Use and hosted browser infrastructure such as Browserbase, but the developer entry point is CLI-first and evidence-oriented.
For comparison positioning, read Tangle Browser Agent vs Browserbase and Browser Use. For wallet safety, read natural-language E2E testing for wallet apps.
Evidence Bar
A useful browser-agent run should leave enough behind that a reviewer can reject it without rerunning the task:
| Artifact | What it answers |
|---|---|
| screenshot | what did the user-visible page show? |
| DOM excerpt | what did the page expose structurally? |
| action log | what did the agent do and in what order? |
| stop reason | why did the agent stop here? |
| error context | was the failure the site, model, network, or instruction? |
Without that evidence, browser automation becomes anecdotal. With it, the run can feed QA, design review, accessibility review, and regression testing.
What This Does Not Prove
A browser run does not prove business correctness. It proves observed browser state for a task. A good run should stop before destructive actions, preserve screenshots, record DOM evidence, and make the final condition explicit.
Start
Install the scoped package, run bad run --help, then run one read-only browser task against a public page. Use that trace as the baseline before moving to authenticated or payment flows.
FAQ
What is browser automation for AI agents?
It is goal-driven browser control where an agent uses DOM state, screenshots, and recovery loops to complete and verify web tasks.
Is Tangle Browser Agent a replacement for Playwright?
No. It uses browser automation concepts around Playwright-style control, but adds an agent loop, CLI workflow, and evidence artifacts.
What should a browser agent capture?
At minimum it should capture the goal, final state, screenshot evidence, DOM evidence, errors, and the stop reason.