Adaptive Learning Pathways

Overview

Sprout adapts your learning path in real-time based on diagnostic assessments and ongoing performance. Unlike static courses that force everyone through the same content, Sprout’s AI agents observe your strengths and gaps, then restructure the knowledge graph to match your needs.

Adaptive learning happens at multiple levels: subconcepts, concepts, and entire topic paths can be modified based on your performance.

Diagnostic Assessment System

Pre-Concept Diagnostics

Before teaching any concept, Sprout runs a diagnostic assessment to understand your current knowledge:

Question Generation

The Subconcept Bootstrap Agent creates 5-10 diagnostic questions per concept—mixed multiple-choice and open-ended.

Difficulty Variation

Questions span easy to challenging to accurately map your understanding across the concept spectrum.

Single Pass

Diagnostics are one-time assessments (no repeats) to minimize testing overhead while gathering key insights.

Optional Completion

You can skip diagnostics entirely, but answering them unlocks powerful personalization.

Question Formats

Multiple Choice
Open-Ended

Format: mcq

4 answer options
Single correct answer
Instant feedback after submission
Stored as { selectedOption: string } in the database

{
  "format": "mcq",
  "prompt": "What is the time complexity of binary search?",
  "options": ["O(n)", "O(log n)", "O(n²)", "O(1)"],
  "correctAnswer": "O(log n)"
}

Format: open_ended

Free-text response
Graded by the Grade Answers Agent (not simple string matching)
Receives score (0.0-1.0) and detailed feedback
Tests deeper understanding vs. recognition

{
  "format": "open_ended",
  "prompt": "Explain why quicksort has O(n²) worst case.",
  "correctAnswer": null  // Graded semantically, not literally
}

Grading Process

When you submit diagnostic answers, the Concept Refinement Agent triggers the Grade Answers Agent:

Extract answers: Latest answer per question
Call grading sub-agent: Claude evaluates each response
Store results: isCorrect (boolean), score (0.0-1.0), feedback (string)
Update database: Graded answers persist for downstream analysis

// From concept-agent.ts
const graded = await gradeAnswers(conceptNode.title, gradableQuestions);

for (const result of graded) {
  await db.update(answers)
    .set({
      isCorrect: result.isCorrect,
      score: result.score,
      feedback: result.feedback,
    })
    .where(eq(answers.id, latestAnswer.id));
}

Graph Adaptation (ORAV Loop)

After grading, the Concept Refinement Agent enters its Observe-Reason-Act-Verify loop:

1. Observe Phase

Current Graph
Student History
Diagnostic Results

The agent calls get_current_subconcepts to view the existing subconcept DAG:

{
  "subconcepts": [
    { "id": "...", "title": "Binary Search Trees", "prerequisites": [] },
    { "id": "...", "title": "Tree Traversal", "prerequisites": ["..."] }
  ]
}

The agent calls check_student_history to see cross-concept performance:

Mastery scores for completed subconcepts
Number of completed nodes
Overall difficulty level
Patterns of struggle or ease

2. Reason Phase

Claude analyzes the data to identify:

Mastered Content

Subconcepts you already understand based on diagnostic performance. These can be removed to save time.

Knowledge Gaps

Missing prerequisites indicated by wrong answers. Requires adding bridge subconcepts.

Misconceptions

Specific incorrect beliefs (e.g., confusing pass-by-value vs. pass-by-reference). Needs targeted remediation.

3. Act Phase

The agent has powerful tools to restructure your learning path:

Add Subconcepts
Remove Subconcepts
Insert Concepts

Tool: add_subconceptCreates new subconcept nodes for:

Bridge topics that fill prerequisite gaps
Remedial content for misconceptions
Enrichment for advanced learners

{
  title: "Pass-by-Reference Fundamentals",
  description: "Clarify memory vs. value semantics",
  prerequisiteIds: ["variables_concept_id"],
  insertBefore: "pointers_subconcept_id"
}

Tool: remove_subconceptDeletes mastered subconcepts and automatically reconnects edges:

Finds all incoming edges to the removed node
Finds all outgoing edges
Creates direct connections from prerequisites to dependents
Validates DAG structure (no cycles)

The agent can remove multiple subconcepts in one adaptation cycle.

Tool: add_prerequisite_concept / add_followup_conceptAdds entire concepts to the topic path:

Prerequisite: Insert before current concept when gaps are too large
Follow-up: Insert after current concept for reinforcement/enrichment

This is a major structural change—the topic path itself is reordered.

4. Verify Phase

Tool: validate_graph After making changes, the agent calls validation to ensure graph integrity:

Orphan detection: All nodes must be reachable from the root
Broken edges: Source and target nodes must exist
Cycle detection: DAG structure must be preserved
BFS traversal: Confirms every node is in the connected component

If issues are found, the agent fixes them and verifies again before completing.

The double-verify pattern ensures no malformed graphs are ever committed to the database.

Mastery Tracking

Subconcept Mastery

As you work through tutoring sessions, the Tutor Agent tracks mastery at the subconcept level:

// From tutor-chat.ts
await db.insert(userNodeProgress).values({
  userId,
  nodeId: subconceptNodeId,
  masteryScore: calculatedScore,  // 0.0 to 1.0
  completed: masteryScore >= 0.7,
  updatedAt: new Date(),
});

Mastery thresholds:

0.0 - 0.4: Struggling (tutor provides more examples)
0.5 - 0.6: Developing (tutor asks follow-up questions)
0.7 - 1.0: Mastered (tutor marks complete, unlocks next nodes)

Exercise Recording

Tool: record_exercise_result After each tutoring exercise:

Student’s answer is stored
Correctness is evaluated
Mastery score is updated (running average)
Progress record is updated in userNodeProgress table

Prerequisite Checking

Tool: check_prerequisite_mastery Before teaching a subconcept, the tutor verifies:

All prerequisite subconcepts are marked completed: true
Mastery scores meet threshold (≥0.7)
Edges in the graph match actual readiness

If prerequisites aren’t met, the tutor redirects you to complete them first.

Personalization Examples

Example 1: Mastered Content Removal

Scenario: You ace diagnostic questions about basic loops. Adaptation:

Concept Refinement Agent sees 100% correctness on loop questions
Reasons that “For Loops” and “While Loops” subconcepts are redundant
Removes both subconcepts from the graph
Reconnects edges to “Nested Loops” (next topic)
Validates graph integrity
Result: You skip 2 subconcepts and save ~30 minutes

Example 2: Knowledge Gap Insertion

Scenario: You miss questions about variable scope. Adaptation:

Concept Refinement Agent sees 0% correctness on scope questions
Reasons that “Lexical Scoping” is a missing prerequisite
Calls add_subconcept to insert “Lexical Scoping Fundamentals”
Wires it as prerequisite to “Closures” subconcept
Validates graph
Result: You learn the foundation before tackling advanced topics

Example 3: Concept-Level Adjustment

Scenario: Diagnostic reveals you lack fundamental algebra skills. Adaptation:

Concept Refinement Agent sees widespread struggles across calculus diagnostics
Reasons that entire algebra prerequisite concept is missing
Calls add_prerequisite_concept to insert “Algebraic Foundations”
Repositions current concept to depend on new prerequisite
Validates graph
Result: A new concept node appears in your topic path

Continuous Adaptation

Adaptation doesn’t stop after diagnostics:

Mid-Tutoring Adjustments

If you struggle during tutoring sessions, the Tutor Agent can request on-the-fly graph changes (feature in development).

Post-Completion Review

After completing a concept, the Review Learning Path Agent suggests enrichment topics or remediation if patterns suggest gaps.

Cross-Concept Learning

The system tracks performance across multiple concepts, identifying global strengths and weaknesses.

Difficulty Calibration

Exercise difficulty adapts based on recent mastery scores—harder problems appear when you’re excelling.

Technical Implementation

Database Schema

Adaptive learning relies on persistent state:

// User progress per node
userNodeProgress: {
  userId: string,
  nodeId: string,
  masteryScore: number,     // 0.0 to 1.0
  completed: boolean,
  exercisesAttempted: number,
  updatedAt: Date
}

// Diagnostic answers
answers: {
  id: string,
  userId: string,
  questionId: string,
  answerData: JSON,         // { selectedOption } or { text }
  isCorrect: boolean,
  score: number,            // 0.0 to 1.0
  feedback: string,
  submittedAt: Date
}

Agent Communication

Agents share data via tool results:

// Concept Refinement Agent calls Grade Answers Agent
const graded = await gradeAnswers(conceptTitle, questions);

// Grade Answers Agent returns structured data
return [
  {
    questionId: "...",
    isCorrect: false,
    score: 0.3,
    feedback: "Your answer confused pass-by-value with pass-by-reference..."
  }
];

This enables complex multi-agent workflows without tight coupling.

Benefits of Adaptive Learning

Time Efficiency

Skip content you’ve already mastered. Focus only on knowledge gaps.

Reduced Frustration

Avoid being thrown into advanced topics without proper prerequisites.

Personalized Pacing

Fast learners progress quickly; those who need more time get additional support.

Targeted Remediation

Address specific misconceptions instead of reviewing entire topics.

Enrichment Opportunities

High performers unlock advanced content and enrichment topics.

Transparent Logic

Watch agent reasoning in real-time via SSE events—understand why changes happen.

Adaptive learning is the core differentiator that makes Sprout an AI tutor, not just an AI content generator.

Get Started

Core Features

Learning Experience

Architecture

​Overview

​Diagnostic Assessment System

​Pre-Concept Diagnostics

Question Generation

Difficulty Variation

Single Pass

Optional Completion

​Question Formats

​Grading Process

​Graph Adaptation (ORAV Loop)

​1. Observe Phase

​2. Reason Phase

Mastered Content

Knowledge Gaps

Misconceptions

​3. Act Phase

​4. Verify Phase

​Mastery Tracking

​Subconcept Mastery

​Exercise Recording

​Prerequisite Checking

​Personalization Examples

​Example 1: Mastered Content Removal

​Example 2: Knowledge Gap Insertion

​Example 3: Concept-Level Adjustment

​Continuous Adaptation

Mid-Tutoring Adjustments

Post-Completion Review

Cross-Concept Learning

Difficulty Calibration

​Technical Implementation

​Database Schema

​Agent Communication

​Benefits of Adaptive Learning

Time Efficiency

Reduced Frustration

Personalized Pacing

Targeted Remediation

Enrichment Opportunities

Transparent Logic

Build docs developers (and LLMs) love

Overview

Diagnostic Assessment System

Pre-Concept Diagnostics

Question Formats

Grading Process

Graph Adaptation (ORAV Loop)

1. Observe Phase

2. Reason Phase

3. Act Phase

4. Verify Phase

Mastery Tracking

Subconcept Mastery

Exercise Recording

Prerequisite Checking

Personalization Examples

Example 1: Mastered Content Removal

Example 2: Knowledge Gap Insertion

Example 3: Concept-Level Adjustment

Continuous Adaptation

Technical Implementation

Database Schema

Agent Communication

Benefits of Adaptive Learning