Prompteryx Browser Agent

AI Node

AI-powered browser automation with computer vision

Overview

The Prompteryx Browser Agent is an AI-powered node that can control the browser using computer vision. Unlike traditional nodes that rely on CSS selectors, this agent "sees" the page and can interact with elements based on visual context and natural language instructions.

When to Use This Node

Use the Prompteryx Browser Agent when:

Traditional selectors are unreliable or constantly changing
The page has complex dynamic content (React, Angular, etc.)
You need reasoning to complete a task (e.g., "find the best deal")
Elements are in iframes, shadow DOM, or rendered in canvas
You want to automate without writing any selectors

Visual Studio with the Prompteryx Browser Agent node on the canvas after a Manual Trigger, and its config panel open showing Instructions, Model and the optional Structured Output Schema — The node in Visual Studio — add it from the Node Library's AI category; the config panel takes plain-language Instructions, a Model, and an optional structured-output schema.

Parameters

Configuration options for this node

Instructions

Text Area

Required

Natural language instruction for what the agent should do

Example: Click the 'Add to Cart' button for the blue shirt

Model

Select

Which model drives the agent (see Model & Credit Cap below)

Default: Prompteryx Core

AI Vision

Select

Required

How much visual detail the agent uses to read each page, from Ultra Saver to Max Precision. Higher = more accurate but uses more credits

Default: Low

Max Steps

Number

Required

Maximum number of actions the agent can take to complete the task (1–50)

Default: 30

Output Variable

Text

Required

Stores the agent's final response for later nodes

Default: agentOutput

How It Works

1. Takes a Screenshot

The agent captures a screenshot of the current page state and sends it to the AI model along with your prompt.

2. AI Analyzes & Decides

The AI model analyzes the screenshot, understands the page structure, and determines what action to take to accomplish your goal.

3. Executes Action

The agent executes the action (click, type, scroll, etc.) based on the AI's decision. It then checks if the task is complete or needs more steps.

Usage Examples

Example 1: Find and Click a Button

No selectors needed - just describe what you want

Prompt:Click the "Sign In" button in the top right corner

Example 2: Fill a Form

Complete multi-step form filling

Prompt:

Fill in the contact form with: Name "John Doe", Email "john@example.com", Message "Hello, I'm interested in your services", then click Submit

Example 3: Decision-Based Task

Let AI make decisions based on page content

Prompt:Find the product with the lowest price that has at least 4 stars rating, then add it to the shopping cart

Example 4: Navigate Complex UI

Handle multi-step navigation

Prompt:Go to Settings, then Privacy & Security, then scroll down and click "Download my data"

Converting to Workflow Nodes

To turn an agent run into traditional Visual Studio nodes, run the task on the AI Browser Agent page and click Refine in Visual Studio (or Open in Visual Studio on the run's "All done" card). This allows you to:

Run the same automation without using AI credits
Fine-tune individual steps for better reliability
Add error handling and conditional logic
Share the workflow with team members

Use the agent once to "teach" it the process, then convert to nodes and replay without AI costs - ideal for production-scale, repeatable work.

AI Vision Settings

The node's AI Vision field offers nine presets, from Ultra Saver to Max Precision (new nodes start on Low):

Ultra Saver · Low · Saver

Less detail, lowest credit usage. Best for simple pages with clear, large elements.

Efficient · Balanced · Detailed

Balanced quality and cost. Works well for most websites and applications.

Enhanced · Precision · Max Precision

The most detail for maximum accuracy. Use for complex pages with small text or detailed interfaces. Higher credit usage.

Final Step Vision

Most agent steps are navigation — clicking, typing, scrolling — which work fine at a low (cheap) vision quality. But the final step is often where the agent has to read or summarise information from the page (a date, a value, a block of text), and that can need more detail.

Final Step Vision lets you set a different quality preset for that final read only — typically higher for accurate extraction, but it can also be lower to save cost. It is optional: leave it on “Same”and the agent uses your normal AI Vision for every step with no change.

How it works

When set to anything other than “Same”, after the agent finishes navigating it does one extra read of the final page at the chosen resolution and produces its answer from that clearer (or cheaper) view. So your cheap navigation/click steps are unaffected — only the final summarise/extract read uses this resolution. The extra read counts toward your credit usage and the per-execution credit cap.

Model & Credit Cap

Model

Choose which model drives the agent — Prompteryx Core (the default for new nodes; cheaper and faster than Prime, strong on forms), Prompteryx Prime(our best-tested automation stack), Prompteryx Swift (fastest and cheapest), Prompteryx Prime+ (the Prime stack on a newer brain), Gemini 4 Argon(Google’s most capable computer-use model), or one of Claude Fable 5.1, Claude Fable 5, Claude Opus 5, GPT-6 Astra, GPT-5.6 Sol and GPT-5.6 Terra. The Claude and GPT models run on our cloud agent runtime, so use them on server-side runs (scheduled, API and Cloud Browser executions) — not a local Chrome run. Admin accounts also see the raw Gemini models. (Prompteryx Jev is available on the AI Browser Agent page only.)

Max Credits (this execution)

A hard cap on the AI credits this single agent run may spend. If the run exceeds it, the node stops with a clear error instead of burning credits in a stuck loop. Set to 0 for no limit.

Run-dialog parity dials

The node also carries the AI Browser Agent’s execution dials: Thinking Level, Adaptive Batching (on by default for new nodes), Fresh-Session Reset (with its every-N-turns threshold), and an optional Structured Output Schema — when a JSON Schema is set, the run also stores JSON matching it in {Output Variable}Json alongside the written answer. It also has Use Residential Proxy (cloud/server runs only; uses metered proxy data) and Safety Decisions (Auto-Consent by default, for unattended runs). (The Show Agent Thinking toggle was removed from this node in Sep 2026; workflows that saved it keep running unchanged.)

Parallel Agents (cloud runs)

Set Parallel Agents above 1 to fan the task out across N cloud agents on server-side executions (scheduled, API, and cloud-browser runs) — each agent takes a slice of the work (an attached row table splits row-by-row) and the results are merged into the Output Variable. Up to 12, and capped by your plan’s limit. A local Chrome run always uses a single agent.

AI Credits & Pricing

Credit Usage

The Prompteryx Browser Agent uses AI credits for each action. Credit usage depends on:

AI Vision setting (higher quality = more credits)
The model you choose
Number of steps taken to complete the task
Complexity of the AI's response

Monitor your credit usage in Settings → Usage.

Best Practices

Be Specific in Your Prompts

Clear, specific prompts get better results. Instead of "click the button", say "click the blue 'Submit' button at the bottom of the form".

Set Reasonable Max Steps

New nodes allow 30 steps (up to 50). For a short task, lower it — e.g. 10 steps for a 3-5 action task leaves room for retries. Too many steps wastes credits on a stuck run; too few may not complete the task.

Use for Teaching, Run with Nodes

Use the AI Browser Agent to learn a new automation, convert it to nodes with Refine in Visual Studio, then run the nodes for production. This minimizes ongoing AI credit costs.

Not for High-Frequency Tasks

For tasks running hundreds of times per day, convert to traditional nodes. The agent is best for exploration, complex one-offs, and creating workflows.

Related Documentation

AI Browser Agent

Full AI browsing experience

Browser Nodes

Traditional selector-based automation

Selectors Guide

When selectors work well

Testing Workflows

Debug AI agent actions

Overview

When to Use This Node

Use the Prompteryx Browser Agent when:

Traditional selectors are unreliable or constantly changing
The page has complex dynamic content (React, Angular, etc.)
You need reasoning to complete a task (e.g., "find the best deal")
Elements are in iframes, shadow DOM, or rendered in canvas
You want to automate without writing any selectors

The node in Visual Studio — add it from the Node Library's AI category; the config panel takes plain-language Instructions, a Model, and an optional structured-output schema.

How It Works

1. Takes a Screenshot

The agent captures a screenshot of the current page state and sends it to the AI model along with your prompt.

2. AI Analyzes & Decides

The AI model analyzes the screenshot, understands the page structure, and determines what action to take to accomplish your goal.

3. Executes Action

The agent executes the action (click, type, scroll, etc.) based on the AI's decision. It then checks if the task is complete or needs more steps.

Usage Examples

Example 1: Find and Click a Button

No selectors needed - just describe what you want

Prompt:Click the "Sign In" button in the top right corner

Example 2: Fill a Form

Complete multi-step form filling

Prompt:

Fill in the contact form with: Name "John Doe", Email "john@example.com", Message "Hello, I'm interested in your services", then click Submit

Example 3: Decision-Based Task

Let AI make decisions based on page content

Prompt:Find the product with the lowest price that has at least 4 stars rating, then add it to the shopping cart

Example 4: Navigate Complex UI

Handle multi-step navigation

Prompt:Go to Settings, then Privacy & Security, then scroll down and click "Download my data"

Converting to Workflow Nodes

Run the same automation without using AI credits
Fine-tune individual steps for better reliability
Add error handling and conditional logic
Share the workflow with team members

Use the agent once to "teach" it the process, then convert to nodes and replay without AI costs - ideal for production-scale, repeatable work.

AI Vision Settings

The node's AI Vision field offers nine presets, from Ultra Saver to Max Precision (new nodes start on Low):

Ultra Saver · Low · Saver

Less detail, lowest credit usage. Best for simple pages with clear, large elements.

Efficient · Balanced · Detailed

Balanced quality and cost. Works well for most websites and applications.

Enhanced · Precision · Max Precision

The most detail for maximum accuracy. Use for complex pages with small text or detailed interfaces. Higher credit usage.

Final Step Vision

How it works

Model & Credit Cap

Model

Max Credits (this execution)

A hard cap on the AI credits this single agent run may spend. If the run exceeds it, the node stops with a clear error instead of burning credits in a stuck loop. Set to 0 for no limit.

Run-dialog parity dials

Parallel Agents (cloud runs)

Best Practices

Be Specific in Your Prompts

Clear, specific prompts get better results. Instead of "click the button", say "click the blue 'Submit' button at the bottom of the form".

Set Reasonable Max Steps

Use for Teaching, Run with Nodes

Use the AI Browser Agent to learn a new automation, convert it to nodes with Refine in Visual Studio, then run the nodes for production. This minimizes ongoing AI credit costs.

Not for High-Frequency Tasks

For tasks running hundreds of times per day, convert to traditional nodes. The agent is best for exploration, complex one-offs, and creating workflows.