Curriculum
3 Tracks, 12 Paths. Frequent milestones, branching choices, daily quests.
3
Tracks
12
Paths
~40h
Total
6,000
Bonus XP
~5d
Avg Completion
~40h includes lessons, exercises, flashcard review, and reflection time
Track 1 · Awakening
L1 Explorer → L2 Apprentice
You discover the power of agents. Master the core primitives: what an agent is, how tools work, and how to control LLM output. Everything else builds on this.
P1
Agent Basics
+500 XP
SDK query() for await..of message.type session_id
What is an Agent
LLM ≠ Agent; Loop, State, Tools, Termination. query() as the simplest agent.
+50
The Minimal Agent Loop
Input → Reason → Act → Output. for await streaming maps message types to loop stages.
+50
Failure Taxonomy
Hallucination, tool misuse, unverifiable output. error message handling.
+50
Turn a chatbot into a stateful loop
Capture session_id, resume conversation, observe state persistence.
+100
Gate: run query(), capture session_id, explain loop vs one-shot call
P2
Tool Use
+500 XP
SDK allowedTools disallowedTools tools config canUseTool
Function Calling Mental Model
Schema → Call → Result → Continue. Observe tool_call and tool_result in stream.
+50
Tool Design Principles
Atomicity, clear descriptions, helpful errors
+50
Tool Orchestration
When to call, when not to, result into context
+50
Implement Calculator Tool
allowedTools: ["Bash"] agent uses node -e for computation
+100
File & Search Tool Mocks
Parameter validation + error handling
+100
Build a canUseTool Gate
Permission callback: log, filter, control via canUseTool
+100
Gate: tool calls logged, canUseTool blocks dangerous ops, multi-tool sequences observed
P3
Prompts & Structured Output
+500 XP
SDK systemPrompt outputFormat Zod schema structured_output maxThinkingTokens
System Prompt Boundaries
Behavioral constraints vs task instructions
+50
Structured Output with Schemas
Pydantic / Zod → JSON Schema → type-safe output. outputFormat + structured_output
+50
Reducing Hallucination
Retrieval, constraints, verification
+50
Prompt ↔ Tool Interaction
"Only output executable content" strategy
+50
Extended Thinking & Reasoning Control
Adaptive thinking, effort parameter. maxThinkingTokens cost/quality tradeoff.
+50
Schema-constrained LLM Output
Define model, validate 99% parseable
+100
Compare 2 Prompt Strategies
Generic vs tool-dense, measure failure rate
+100
Gate: structured output validates 100%, system prompt preset understood, thinking budget tradeoff measured
Reflection Day
Pause. Review what you've built. Summarize agent primitives in your own words. Compress before expanding.
+50 XP
Track 2 · Forging
L2 Apprentice → L3 Builder
You forge your own tools. Memory, context engineering, evaluation, model routing, design patterns, and your first complete agent framework. Completing this track means you can build agent loops independently.
P4
Memory & Context Engineering
+500 XP
SDK resume forkSession continue v2 Session APIs betas: context-1m
Short-term vs Long-term Memory
Conversation summary, vectors, events. Short-term = session; Long-term = resume.
+50
Write Strategy
What to write, when, how to compress
+50
Memory Corruption & Repair
Stale info, wrong facts, injection attacks
+50
Context Window Management
Sliding window, threshold compression, prefix caching (92% reuse). betas: context-1m
+50
Prefix Reuse & Cost Optimization
Cache-friendly prompt structure, long-context pricing, batching
+50
Session Summarizer
Key field extraction: constraints, preferences
+100
Task Memory Cards
Task → decision → outcome → lesson learned
+100
Build a Context Compressor
Auto-compress at threshold, preserve key info, measure token savings
+100
Gate: resume, fork, continue sessions demonstrated. V2 vs V1 session model explained.
P5
Evaluation, Routing & Recovery
+500 XP
SDK fallbackModel setModel() supportedModels() error codes accountInfo()
Why Agents Must Be Evaluated
No eval = no iteration
+50
Eval Types Spectrum
Parse → typecheck → unit → golden → LLM-as-judge
+50
Statistical View: N Trials
Agent is stochastic — single pass ≠ reliable
+50
Cost-Aware Evaluation
Token budget vs success rate tradeoff
+50
Error Recovery Strategies
Retry, minimal fix, rollback, degrade. fallbackModel for auto-recovery.
+50
Model Selection Strategy
Opus/Sonnet/Haiku tradeoffs, cost vs capability matrix
+50
Intelligent Model Routing
Small model classifies → big model reasons, dynamic escalation via setModel()
+50
Output Parser + Schema Validator
Build a grader that checks structure
+100
Run 20 Trials, Compute Stats
Success rate, variance, avg token cost
+100
Generate → Execute → Verify → Fix
Error fed back until pass or max retries
+100
Build a Cost-Aware Model Router
Route by complexity, compare cost/quality across models via supportedModels()
+100
Gate: 20-trial eval report, model router picks optimal model, cost tracked per route
P6
Design Patterns & Capstone
+500 XP
SDK hooks.PreToolUse hooks.PostToolUse hooks.Stop settingSources permissionMode: plan
5 Workflow Patterns
Chaining, Routing, Parallelization, Orchestrator-Worker, Evaluator-Optimizer
+50
When NOT to Use an Agent
If-else → Workflow → Agent decision tree
+50
Observability
Structured logging, tracing, cost tracking via hooks.PreToolUse + hooks.PostToolUse
+50
Agent Bootstrap & Initialization
Cold start vs warm start, repo warmup, permissionMode: plan for planning-before-execution
+50
Routing vs Orchestrator-Worker
Implement both, compare tradeoffs
+100
Capstone: Mini Agent Framework
Tool registry + loop + eval + fix + logging — 7+ SDK features, 5 tests ≥ 4/5 pass
+500
Gate: L3 Builder unlocked — capstone uses hooks, structured output, fallback, sessions, settings
Reflection Day
You can now build agents independently. Review your capstone. Choose your expedition direction below.
+50 XP
Track 3 · Expedition
L3 Builder → L5 Expert
You set out into uncharted territory. Deep specializations — choose your own adventure and complete any 2 paths to unlock the Production capstone.
Choose 2+ paths below. Different choices = different Badge Shelves.
Recommended Routes
Full-Stack Agent
Build & ship complete agent systems
P7P10P12
AI Engineer
Knowledge systems & multi-agent orchestration
P8P9P12
Safety-First
Trust, control & production readiness
P7P11P12
P7
MCP & Tool Ecosystem
+500 XP
SDK createSdkMcpServer() tool() mcpServers mcp__*__* mcpServerStatus()
MCP Protocol Structure
Server / client / tool / resource. createSdkMcpServer() + tool() in-process.
+50
Using Existing MCP Servers
GitHub, filesystem, database integrations via mcpServers stdio/http config.
+50
Build a Minimal MCP Server
Expose custom tools via mcp__*__* naming, verify with mcpServerStatus()
+100
Gate: MCP server runs, agent calls mcp__*__* tools, mcpServerStatus() connected
P8
Agentic RAG
+500 XP
SDK MCP + subagents outputFormat for relevance scoring multi-step tool calls
Traditional RAG vs Agentic RAG
Agent decides when, what, whether to re-retrieve. MCP search tool + outputFormat relevance scoring.
+50
Multi-step Retrieval Strategy
Decompose → retrieve → evaluate → re-retrieve
+50
Self-deciding Search Agent
Judge if retrieval needed or answer directly
+100
Iterative RAG with Quality Check
Retrieve → check relevance → re-query if insufficient
+100
Gate: multi-step retrieval, relevance scores above threshold, re-query count tracked
P9
Multi-Agent Systems
+500 XP
SDK agents{} AgentDefinition model: haiku|sonnet|opus maxTurns hooks.SubagentStart/Stop
When to Go Multi-Agent
Specialization, parallelism, separation of concerns. agents{} with model overrides.
+50
Orchestration Patterns
Orchestrator-Worker, Peer-to-Peer, Pipeline. maxTurns + model per agent.
+50
Communication & State Sharing
Message passing, shared memory, handoff
+50
Planner + Executor Duo
One agent plans, another executes tools
+100
Writer + Reviewer Pipeline
Generate → critique → revise cycle
+100
Gate: 2+ agents run, SubagentStart/Stop hooks fire, cleanup implemented
P10
Code Generation Agent
+500 XP
SDK Read/Write/Edit/Bash outputFormat for plans enableFileCheckpointing rewindFiles()
Template-based Generation
Fill template > full text; DSL + strict interface
+50
Compile / Typecheck / Render Tools
Layered verification for generated code
+50
Self-healing Strategies
Compile → runtime → output assertion, prioritized. enableFileCheckpointing + rewindFiles() for rollback.
+50
Strict Scene Schema
Schema → compiler → LLM fills schema only
+100
Typecheck + Render Preview
Structured error output + low-res preview
+100
3-layer Self-healing Loop
TS error → runtime exception → visual assertion
+100
Gate: 10 scripts, ≥ 70% success, repair logs with diffs, file checkpointing demonstrated
P11
Human-in-the-Loop Design
+500 XP
SDK permissionMode (all 4) canUseTool (advanced) AskUserQuestion hooks.PermissionRequest
Approval Flows & Permission Tiers
When to auto-execute vs ask. permissionMode: default → acceptEdits → bypass → plan.
+50
Checkpoint & Rollback
Undo-safe execution, AskUserQuestion for confirmation, audit trails.
+50
Calibrating Autonomy
Progressive trust: fully supervised → semi-auto → full auto
+50
Build a Permission-Gated Agent
Classify actions by risk via canUseTool, confirm destructive ops, log decisions.
+100
Gate: 3-tier permission system works, destructive ops blocked, AskUserQuestion demonstrated
P12
Production & Final Capstone
+500 XP
SDK sandbox{} all 12 hooks plugins[] known issues (14)
Sandboxing & Safety
Isolated execution, sandbox config, network control, command exclusion.
+50
Rate Limiting & Cost Control
Token budgets, retry budgets, circuit breakers via hooks control plane.
+50
Monitoring & Alerting
Success rate dashboards, anomaly detection
+50
Add Production Guardrails
Token budget + timeout + structured logging
+100
Final Capstone: End-to-End Agent
MCP tools + subagents + eval harness + sandbox + all 12 hooks + N-trial report + public repo
+500
Gate: reproducible demo, eval report, cost report, sandbox verified — L5 Expert unlocked
Graduation Reflection
You've completed the expedition. Review your capstone, reflect on your journey, and plan what's next.
+50 XP
Progression
Level ↔ Path Map
L1
Explorer — 0 XP
Start P1 Agent Basics
L2
Apprentice — 1,500 XP
Complete P1+P2, working on P3
L3
Builder — 6,000 XP
Complete Track 1+2 through P6 Capstone
L4
Architect — 8,000 XP
Complete 2+ Advanced paths
L5
Expert — 10,000 XP
Complete P12 Production + Final Capstone
L6
Master — 20,000 XP
Create courses, MCP servers, open-source
Economics
XP Economy
Path XP Breakdown
Each path awards 500 XP total = content XP (lessons + exercises) + completion bonus. The 500 XP is granted as a one-time reward upon path completion, not subject to daily cap.
Metric
Value
Path completion rewards
12 × 500 = 6,000 XP
Daily cap (regular activities)
300 XP
Daily quest XP
+20–30 XP
Reflection day bonus
3 × 50 = 150 XP
First Path complete
~Day 3
Avg completion interval
~5 days
Branch choices (Autonomy)
5 options, pick 2+
Days to L3 Builder (6,000 XP)
~20 days
Days to L5 Expert (10,000 XP)
~35 days
Streak protection
1 rest day/week
Streak protection: 1 rest day per week won't break your streak. Sustainable learning over grinding.
Achievements
Badge Triggers
Knowledge
Initiated1st lesson
Versed10 lessons
ScholarAll lessons (40+)
Practice
Tried1st exercise
Skilled10 exercises
MasterAll exercises (25+)
Habit
Started3-day streak
Consistent7-day streak
Unstoppable30-day streak
Social badges (Reviewer, Mentor, Pillar) are in the Special Badges category.
Micro-engagement
Daily Quests
5-minute rituals to keep the streak alive. One quest per day — lower the entry barrier, raise the DAU.
Quick Prompt
Given a scenario, write the system prompt or pick the right tool
+20 XP
Failure Spot
Read an agent trace, find the failure point and classify the error type
+20 XP
Tool Match
Given a task description, select the correct tool + parameters from a set
+20 XP
Code Review
Review another learner's exercise submission, leave constructive feedback
+30 XP
Code Review Flow: Peer-to-peer review with AI-assisted quality checks. Submit your exercise → get matched with another learner's submission → leave feedback → earn XP when your feedback is marked helpful.
Validation
Design Rationale
Frequent milestones
Gamification meta-analysis: "Rules/Goals + Challenge" yields highest impact (d=0.566). 12 path completions at ~5-day intervals vs 2 at ~25-day intervals.
Autonomy via branching
SDT research: Autonomy effect g=0.638. Track 3's "choose 2 of 5" with recommended routes provides guided choice without decision fatigue.
Streak Freeze for sustainability
1 rest day/week without breaking streak. Aligns with "sustainable growth" principle.
Zen breathing rhythm
Reflection Days between tracks create deliberate pauses — compress knowledge before expanding. Mirrors context window management in agent design.
Relatedness via Social track
SDT's third pillar. Daily code reviews + Social badges address the biggest gap in solo-learning gamification: community connection.
Competence perception
Gamification has minimal impact on competence (g=0.277). Mitigated by showing real execution results at every gate — not just XP.
Summary
Before vs After
Before
2 mega-paths (~16–20h each)
After
12 focused paths (~2–6h each)
Before
First milestone at ~Day 21
After
First milestone at ~Day 3
Before
No branching, same order
After
5 branches + recommended routes
Before
1,000 XP from path bonuses
After
6,000 XP from path bonuses
Before
Solo progression only
After
Daily quests + social badges + reflections
Before
No context / model routing content
After
Context engineering + model routing + HITL