Claude-Powered AI Agents

Initial message → Claude → [text + tool_use blocks]
                              ↓
                    emit reasoning (visible to you)
                    execute tools
                    feed results back → Claude
                              ↓
                    repeat until complete

POST /api/agents/topics/:topicNodeId/run
        ↓
┌─ Phase 1: Topic Agent ─────────────────┐
│  Generates 6-10 concepts                │
│  Extracts document context              │
│  Saves nodes + edges via SSE            │
└─────────────────────────────────────────┘
        ↓
┌─ Phase 2: Subconcept Bootstrap (×N) ───┐
│  Runs in parallel (max 3 concurrent)    │
│  Creates diagnostic questions           │
│  Creates subconcept DAG per concept     │
│  Streams SSE for each agent             │
└─────────────────────────────────────────┘

POST /api/agents/concepts/:conceptNodeId/run

Phase 1 (no answers): Return diagnostic questions as JSON
Phase 2 (answers exist): Stream concept refinement via SSE

POST /api/chat/sessions/:sessionId/tutor

Sends student message → AI tutor response with reasoning
Passes full context (userId, nodeIds, sessionId) for DB queries

// Tools save directly to the database
await db.insert(nodes).values({
  id: uuid(),
  title: input.title,
  variant: 'subconcept',
  // ...
});

// The agent sees the tool result, not the return value
return JSON.stringify({ success: true, nodeId });

const textBlocks = response.content.filter(
  (block) => block.type === 'text'
);
if (textBlocks.length > 0) {
  onThinking?.(textBlocks.map(b => b.text).join('\n'));
}

1. User creates topic
         ↓
2. Topic Agent generates concepts
         ↓
3. Subconcept Bootstrap Agents run in parallel
         ↓
4. User answers diagnostic questions
         ↓
5. Concept Refinement Agent personalizes graph
         ↓
6. Tutor Agent teaches subconcepts
         ↓
7. Review Agent suggests enrichment