Skip to main content

​
The AI Agent Problem

AI-powered browsers and autonomous agents are proliferating:
  • Arc AI Browser
  • Browser Use
  • Anthropic’s computer use
  • Custom AI agents for workflows
But they’re failing 40-60% of the time.

​
Three Critical Failure Modes

​
1. Action Ambiguity

<button onclick="handleCheckout()">Proceed</button>
AI agent doesn’t know:
  • Does this navigate to a new page?
  • Does this open a modal?
  • Does this trigger an async operation?
  • Is it safe to click without losing work?
Result: Hesitation, incorrect assumptions, failed workflows

​
2. Identification Failure

<button>Cancel</button>
<button>Save Draft</button>
<button>Back</button>
<button>Submit Order</button>
<button>Continue Shopping</button>
AI agent must guess:
  • Which button completes the user’s intent?
  • Which is the primary vs. secondary action?
  • What’s the semantic meaning of each?
Result: Clicking wrong buttons, broken user flows

​
3. Intent Mapping

<!-- Three different patterns, same text -->
<div onclick="addToCart()">Add to Cart</div>
<a href="/cart/add">Add to Cart</a>
<form action="/cart"><button>Add to Cart</button></form>
AI agent sees:
  • Identical visual labels
  • Completely different interaction patterns
  • No semantic hints about actual behavior
Result: Wrong element selected, unexpected side effects

​
Current “Solutions” Don’t Work

​
Heuristic Guessing

Approach: Parse class names, button text, DOM position
  • ❌ Brittle - breaks with design changes
  • ❌ Unreliable - 40-60% failure rate
  • ❌ Maintenance nightmare - every site is different

​
Computer Vision

Approach: Screenshot analysis, visual element detection
  • ❌ Slow - adds latency to every interaction
  • ❌ Expensive - requires heavy processing
  • ❌ Fragile - fails with dynamic content

​
LLM Reasoning

Approach: Feed page HTML to LLM, let it “figure it out”
  • ❌ Inconsistent - different results each time
  • ❌ Costly - tokens add up fast at scale
  • ❌ Unpredictable - can’t guarantee behavior

​
The WCP Solution

​
Explicit > Implicit

Instead of guessing, websites declare intent: 
<button data-wcp="action:purchase; effect:navigate; primary:true">
  Checkout
</button>
Now AI agents know:
  • ✅ This initiates a purchase
  • ✅ It will navigate to a new page
  • ✅ This is the primary action

​
Simple Syntax

One attribute, semicolon-separated: 
data-wcp="key:value; key:value"
Core vocabulary:
  • Actions: add, remove, search, purchase, submit
  • Effects: navigate, modal, async, download
  • Regions: product, checkout, search, navigation

​
Immediate Benefits

95%+ Success Rate

vs. 40-60% with heuristic guessing

Predictable Behavior

Agents know what will happen before clicking

Faster Execution

No visual processing or heavy LLM reasoning needed

Better UX

Reliable agent interactions = happy users

​
Real-World Impact

​
Before WCP

User: “Add these headphones to my cart” AI Agent:
  1. Scans page for “add” + “cart” text
  2. Finds 3 potential matches:
    • Newsletter signup button (contains “add”)
    • Product link (near “cart” icon)
    • Actual add-to-cart button
  3. Guesses based on class names
  4. 60% chance of failure

​
After WCP

User: “Add these headphones to my cart” AI Agent:
  1. Queries for data-wcp="action:add"
  2. Finds annotated button
  3. Verifies effect:async (no navigation)
  4. Clicks with confidence
  5. 95%+ success rate

​
Why Now?

​
The *CP Ecosystem is Emerging

1

October 2024: MCP

Anthropic launches Model Context Protocol - standardizes how LLMs access external data
2

November 2024: ACP

Stripe announces Agent Communication Protocol - standardizes agent-to-agent payments
3

December 2024: WCP

Web Context Protocol - standardizes how agents interact with web pages
Pattern: Infrastructure protocols for the AI era

​
AI Agents Are Scaling

  • Browser automation tools going mainstream
  • Enterprise adopting AI workflows
  • Consumer AI assistants becoming powerful
  • Autonomous agents handling complex tasks
Without standards, chaos. With WCP, reliability.

​
Who Benefits?

  • Control how AI agents interact with your site
  • Prevent wrong actions (accidental deletions, wrong buttons)
  • Enable AI-powered user flows
  • Future-proof for AI-native web
  • 95%+ success rate vs. 40-60%
  • Faster, cheaper execution
  • Predictable behavior
  • Standardized approach across sites
  • Reliable AI assistance
  • Fewer errors and frustrations
  • Faster task completion
  • Better overall experience
  • Standardized protocol for AI interaction
  • Reduced fragmentation
  • Enables innovation
  • Scales with AI adoption

​
Comparison to Alternatives

ApproachSuccess RateSpeedCostMaintainability
Heuristic Guessing40-60%FastLowPoor (breaks often)
Computer Vision60-70%SlowHighMedium
Heavy LLM Reasoning70-80%SlowVery HighMedium
WCP95%+Very FastVery LowExcellent

​
The Network Effect

The more sites adopt WCP, the more valuable it becomes:
1

Early Adopters

First sites implement WCP → Their AI interactions become reliable
2

Agent Support

AI agents add WCP support → Works better on WCP sites
3

Competitive Advantage

Sites without WCP perform worse → Pressure to adopt
4

Standard Practice

WCP becomes expected, like semantic HTML or OpenGraph

​
Join the Movement

Implement WCP

Add to your site in 5 minutes

Read the Spec

Full technical specification

View Examples

Real-world implementations

Contribute

Help shape the protocol
WCP is open source (CC0). Free to use, extend, and modify without restrictions. No vendor lock-in, no licensing fees.