How AMP Works

The Core Problem

Different users are motivated by different things. Some want step-by-step guidance, others prefer to figure things out themselves. Some respond to achievement-focused language, others to learning-focused language.

Traditional AI agents use the same communication style for everyone, leading to:

• Lower task completion rates
• Frustrated users who don't connect with your agent's style
• Wasted engineering time building personalisation logic

The AMP Solution

AMP automatically learns what motivates each user and provides real-time context to adapt your agent's behaviour.

Example Flow

Let's walk through a real example of AMP in action:

1. User Makes a Request

A user asks your coding agent: "Help me build a login page"

2. Agent Requests Context

const context = await amp.getContext({
  userId: "user_123",
  task: "build a login page",
  complexity: "medium",
});

3. AMP Returns Context

Based on the user's historical behaviour, AMP returns:

{
  "requestId": "req_abc123",
  "suggestedFraming": "micro_task",
  "communicationStyle": "brief_directive",
  "complexity": "break_into_steps",
  "rationale": "User has 85% completion rate with step-by-step guidance",
  "confidence": 0.87
}

4. Agent Adapts Its Response

Your agent uses this context to structure its response:

// Build a system prompt using AMP's recommendations
const systemPrompt = `You are a coding assistant.
Communication: Brief and directive (no fluff).
Task framing: Break this into micro-tasks.
Approach: Give step-by-step guidance.`;

// Generate response
const response = await openai.chat.completions.create({
  model: "gpt-4",
  messages: [
    { role: "system", content: systemPrompt },
    { role: "user", content: "Build a login page" },
  ],
});

5. User Completes the Task

The user follows the steps and successfully builds the login page. Your agent reports this success:

await amp.reportOutcome({
  requestId: context.requestId,
  started: true,
  completed: true,
  timeToStart: 30,      // seconds
  flowState: true,      // user seemed engaged
  satisfaction: 0.9,    // optional feedback score
});

6. AMP Learns

AMP updates the user's profile:

• Reinforces that this user responds well to step-by-step guidance
• Notes they prefer brief, directive communication
• Increases confidence in "micro_task" framing for this user

Motivation Dimensions

AMP analyses user behaviour across multiple dimensions:

The Machine Learning Model

AMP uses a multi-armed bandit approach combined with collaborative filtering:

Initial Phase (Cold Start)

• New users start with population-level defaults
• AMP tries different approaches to gather data
• Learns from similar user patterns (collaborative filtering)

Learning Phase

• Builds individual motivation profile from outcomes
• Increases confidence in effective approaches
• Reduces exploration, increases exploitation

Optimised Phase

• High-confidence recommendations based on proven patterns
• Continues to adapt as user behaviour evolves
• Detects and responds to context shifts

Privacy & Data

AMP is designed with privacy in mind:

• Only behavioural patterns are stored, not task content
• User IDs are hashed and anonymised
• No PII (personally identifiable information) is required
• Data is encrypted at rest and in transit
• Full GDPR and CCPA compliance

Next Steps

• Learn about motivation profiles
• Understand the context API
• Master outcome reporting