🗓️ 21032026 2100

AI SYSTEMS LEARNING ROADMAP

A structured reading and learning plan progressing from LLM fundamentals to building custom AI tool implementations

Phase 1: Foundations

Goal: understand how LLMs work and how they interact with tools

ResourceTypeLink
Andrej Karpathy — "Intro to Large Language Models"VideoYouTube
Anthropic — Tool Use DocumentationDocsdocs.anthropic.com
OpenAI — Function Calling GuideDocsplatform.openai.com

Zettelkasten: llm_tool_use

Phase 2: Agent Patterns

Goal: learn how agents are architected and when to use which pattern

ResourceTypeLink
Lilian Weng — "LLM Powered Autonomous Agents"Bloglilianweng.github.io
Andrew Ng — Agentic Design Patterns (DeepLearning.AI)Videodeeplearning.ai
Anthropic — "Building Effective Agents"Bloganthropic.com

Zettelkasten: agentic_design_patterns, agent_memory_and_state

Phase 3: MCP & Tool Infrastructure

Goal: understand the protocol layer that connects agents to tools at scale

ResourceTypeLink
MCP Official DocumentationDocsmodelcontextprotocol.io
MCP Quickstart — Building a ServerTutorialmodelcontextprotocol.io
MCP Quickstart — Building a ClientTutorialmodelcontextprotocol.io

Zettelkasten: model_context_protocol, custom_mcp_servers

Phase 4: Build Your Own

Goal: hands-on implementation to solidify understanding

ProjectDescriptionStarting Point
Custom MCP ServerExpose an internal API or database via MCP. Start with stdio transport, then add HTTPMCP Server Quickstart
Agent with Tool UseBuild an agent that uses tools to accomplish a multi-step taskAnthropic SDK, LangChain Agents, LlamaIndex Agents
Multi-Agent SystemCombine multiple agents using your custom MCP tools — try coordinator or agent-as-tool patterncoordinator_router_pattern, agent_as_tool_pattern

Phase 5: Advanced Topics

Goal: deeper exploration for production-grade systems

TopicWhat to Explore
RAG (Retrieval-Augmented Generation)Embedding models, vector databases, chunking strategies, hybrid search
Agent Evaluation & ObservabilityTracing agent runs, measuring tool call accuracy, detecting loops/hallucinations
Production ConcernsGuardrails, cost management, latency optimization, rate limiting, error recovery

References