Roadmap

Agentic AI Masterclass

From Python and LLM fundamentals to shipping autonomous, tool-using agents in production.

17stages127topics~36hours

Curated from the best, MDN · Kubernetes · AWS · OWASP · Google SRE & more

Agentic AI engineers are among the most in-demand roles in 2025+. Every product team is building autonomous agents, you need to understand the full stack from architecture to security to be the person who ships them.

The complete path, 2 of 127 topics have lessons here; the other 125 are marked learn anywhere. We won't pretend we cover everything.

Ready to build? See capstone projects

Stage 1 / 17 · 9 topics · 1 lessons

Programming & Engineering Foundations

The software-engineering bedrock every agent builder needs before touching models.

Python ProficiencyIdiomatic Python: typing, dataclasses, comprehensions, generators, decorators. Python tutorial

Async & Concurrencyasyncio, coroutines, tasks, semaphores for concurrent tool calls and I/O. Python docs

Type Hints & PydanticStatic typing and Pydantic models for structured, validated agent I/O. Python docs

Git & Version ControlBranching, PRs, code review, and collaborative workflows. Pro Git book

HTTP & REST APIsRequests, status codes, auth headers, pagination, rate limits. MDN

JSON, YAML & SchemasSerialization formats and JSON Schema that drive tool definitions. YAML spec

Testing & DebuggingStart hererecpytest, mocking external calls, debugging non-deterministic systems.

CLI, Linux & ShellrecFilesystem, processes, env vars, and shell scripting for agent runtimes. The Linux Command Line

TypeScript/Node BasicsoptionalJS/TS fundamentals for the LangChain.js and Vercel AI ecosystems. TypeScript docs

Stage 2 / 17 · 10 topics · 0 lessons

ML & LLM Foundations

How language models work under the hood so you reason about agent behavior, not guess.

ML FundamentalsrecTraining vs inference, supervised learning, loss, generalization, overfitting. Google ML Crash Course

Neural Networks & BackpropoptionalLayers, activations, gradient descent at a conceptual level. 3Blue1Brown

Transformer ArchitectureSelf-attention, positional encoding, why transformers dominate LLMs. Jay Alammar

TokenizationBPE, token counting, and how tokens map to cost and context limits. Hugging Face course

Pretraining & Fine-tuningNext-token prediction, instruction tuning, RLHF/DPO alignment. Hugging Face course

Sampling & DecodingTemperature, top-p, top-k, and how they shape determinism. Hugging Face docs

EmbeddingsVector representations of meaning that power retrieval and memory. Pinecone Learn

Context WindowsContext length, lost-in-the-middle effects, and attention costs. Claude docs

Model LandscaperecFrontier vs open models: GPT, Claude, Gemini, Llama, Mistral, Qwen. Hugging Face

Capabilities & LimitationsHallucination, reasoning limits, knowledge cutoffs, brittleness. Claude docs

Stage 3 / 17 · 9 topics · 0 lessons

Prompt Engineering

Precise control of model behavior, the core skill that makes or breaks an agent.

Prompt AnatomySystem, user, assistant roles and message structure. Prompting Guide

Zero-shot & Few-shotIn-context learning with examples to steer outputs. Prompting Guide

Chain-of-ThoughtEliciting step-by-step reasoning for harder tasks. Prompting Guide

ReAct, ToT & Self-ConsistencyReasoning+acting, tree-of-thought, and majority-vote sampling. Prompting Guide

Structured OutputJSON mode, schema-constrained generation, function-call formatting. Claude docs

System Prompt DesignPersona, constraints, and guardrails for agent identity. Claude docs

Prompt TemplatingrecVariables, partials, and reusable prompt components. LangChain docs

Prompt ChainingrecDecomposing tasks into sequenced, composable prompts. Claude docs

Meta & Self-Refine PromptingoptionalModels critiquing and improving their own outputs. Prompting Guide

Stage 4 / 17 · 8 topics · 0 lessons

Working with LLM APIs

Calling, configuring, and operating model providers programmatically.

Chat Completion APIsOpenAI/Anthropic/Gemini SDKs, messages, and response parsing. Claude docs

Function / Tool CallingNative tool-call APIs that let models invoke functions. Claude docs

Streaming ResponsesToken streaming, SSE, and incremental UI updates. Claude docs

Multimodal InputsrecVision, audio, and document inputs in agent pipelines. Hugging Face docs

Token & Cost ManagementCounting tokens, budgeting, and model-tier selection. Claude docs

Prompt Caching & BatchrecCaching, batch APIs, and latency/cost optimizations. Claude docs

Rate Limits & RetriesBackoff, idempotency, and resilient API client design. Claude docs

Local & Self-Hosted ModelsoptionalOllama, vLLM, and quantization for on-prem inference. Ollama

Stage 5 / 17 · 8 topics · 0 lessons

Tool Use & Function Calling

Giving agents hands, the defining capability of an agentic system.

Tool DefinitionsDescribing tools with names, descriptions, and JSON schemas. Claude docs

Tool Dispatch & ExecutionParsing tool calls, executing, and returning results to the model. Claude docs

Parallel Tool CallsrecConcurrent invocation and result aggregation. Claude docs

Tool Error HandlingValidation, graceful failures, and self-correction loops. Claude docs

Code Execution ToolsSandboxed Python/REPL tools for computation and data tasks. Claude docs

Web Search & BrowsingSearch APIs and headless-browser tools for live information. Claude docs

Wrapping External APIsTurning third-party REST/GraphQL APIs into agent tools. Claude docs

Tool Selection & RoutingrecHelping models pick the right tool from large toolsets. Anthropic Engineering

Stage 6 / 17 · 5 topics · 0 lessons

Model Context Protocol & Standards

The emerging interoperability layer connecting agents to tools and data.

MCP FundamentalsWhat MCP is and why standardized tool/context interfaces matter. MCP docs

Building MCP ServersExposing tools, resources, and prompts via an MCP server. MCP docs

MCP Clients & HostsrecConnecting agents to MCP servers and managing capabilities. MCP docs

Transports & Authrecstdio, HTTP/SSE transports and securing MCP connections. MCP docs

Other Interop StandardsoptionalOpenAPI tool specs and agent-to-agent protocols. MCP docs

Stage 7 / 17 · 9 topics · 0 lessons

Retrieval-Augmented Generation

Grounding agents in external knowledge to fight hallucination and stay current.

RAG ArchitectureThe retrieve-then-generate pattern and when to use it. Prompting Guide

Chunking StrategiesFixed, semantic, and recursive chunking with overlap. Pinecone Learn

Embedding ModelsChoosing and evaluating embedding models for retrieval. MTEB Leaderboard

Vector DatabasesPinecone, Weaviate, Qdrant, pgvector and ANN indexing. Pinecone Learn

Hybrid & Keyword SearchrecCombining BM25 with dense retrieval for precision. Elastic

RerankingrecCross-encoder rerankers to sharpen retrieved context. Cohere docs

Query TransformationrecQuery rewriting, HyDE, and multi-query expansion. LlamaIndex docs

Agentic & Graph RAGAgents that iteratively retrieve; GraphRAG over knowledge graphs. LangChain docs

RAG EvaluationrecFaithfulness, context precision/recall, and answer relevance. Ragas docs

Stage 8 / 17 · 8 topics · 0 lessons

Agent Architectures & Reasoning

The control-flow patterns that turn an LLM into an autonomous agent.

The Agent LoopPerceive-reason-act-observe cycles and termination conditions. Anthropic

ReAct AgentsInterleaving reasoning traces with tool actions. Prompting Guide

Planning & DecompositionTask decomposition, plan-and-execute, and subgoal management. Prompting Guide

Reflection & Self-CritiqueReflexion-style loops where agents evaluate and retry. Prompting Guide

Routing & Orchestrator-WorkerrecRouters and supervisor patterns delegating to specialists. Anthropic

Graphs & State MachinesModeling agent control flow as explicit graphs and states. LangGraph docs

Human-in-the-LoopApproval gates, interrupts, and steering for safety. LangChain docs

Autonomous Goal LoopsoptionalLong-running, self-directed agents (AutoGPT-style) and their risks. Anthropic

Stage 9 / 17 · 6 topics · 0 lessons

Memory & State Management

How agents remember, persist, and manage their working context over time.

Short-Term MemoryConversation history and scratchpad within a context window. LangChain docs

Long-Term MemoryPersistent semantic, episodic, and procedural memory stores. LangChain docs

Context Window ManagementSummarization, truncation, and compaction to fit limits. Anthropic Engineering

State Persistence & CheckpointsDurable state, resumable runs, and crash recovery. LangGraph docs

Memory FrameworksrecMem0, Zep, and patterns for entity and fact memory. Mem0 docs

Context EngineeringDeliberately curating what enters the context for each step. Anthropic Engineering

Stage 10 / 17 · 8 topics · 1 lessons

Agent Frameworks & Orchestration

The production toolkits that let you build agents without reinventing plumbing.

LangChainChains, tools, and the broad LLM-app ecosystem. LangChain docs

LangGraphGraph-based stateful agents with checkpointing and HITL. LangGraph docs

LlamaIndexrecData-centric framework for RAG and agentic workflows. LlamaIndex docs

OpenAI Agents SDKrecHandoffs, guardrails, and tracing in OpenAI's agent runtime.

CrewAIrecRole-based multi-agent crews with task delegation. CrewAI docs

AutoGen / AG2optionalConversational multi-agent orchestration patterns. AutoGen docs

PydanticAI / DSPyoptionalType-safe agents and programmatic prompt optimization. PydanticAI docs

Choosing a FrameworkrecTrade-offs vs building from scratch on raw APIs. LangGraph docs

Stage 11 / 17 · 6 topics · 0 lessons

Multi-Agent Systems

Coordinating multiple specialized agents to solve complex tasks.

Coordination PatternsSupervisor, hierarchical, network, and sequential topologies. LangChain docs

Roles & SpecializationDesigning focused agents with clear responsibilities. CrewAI docs

Inter-Agent CommunicationMessage passing, shared state, and blackboard patterns. LangChain docs

Handoffs & DelegationrecTransferring control and context between agents. LangChain docs

Agent-to-Agent ProtocolsoptionalEmerging A2A standards for cross-system agent interop. Hugging Face Agents

Failure ModesrecError compounding, coordination cost, and when single-agent wins. Anthropic Engineering

Stage 12 / 17 · 7 topics · 0 lessons

Evaluation & Observability

Measuring, tracing, and trusting non-deterministic agent systems.

Eval FundamentalsWhy agents need evals and how they differ from ML metrics. Claude docs

Offline EvaluationGolden datasets, regression suites, and task success rates. LangSmith docs

LLM-as-JudgeModel-graded eval, rubrics, and bias mitigation. Hugging Face Cookbook

Trajectory EvaluationrecJudging the path: tool choice, steps, and efficiency. LangSmith docs

Tracing & ObservabilityLangSmith, Langfuse, OpenTelemetry for agent runs. Langfuse docs

Online Eval & FeedbackrecProduction monitoring, user feedback, and A/B tests. LangSmith docs

Agent BenchmarksrecSWE-bench, GAIA, WebArena, tau-bench and their limits. GAIA Leaderboard

Stage 13 / 17 · 8 topics · 0 lessons

Safety, Security & Alignment

Keeping autonomous, tool-wielding agents from causing harm.

Prompt InjectionDirect and indirect injection attacks and defenses. OWASP GenAI

Guardrails & ValidationInput/output filters, content policies, and schema enforcement. Claude docs

Sandboxing & Least PrivilegeIsolating code execution and scoping tool permissions. Claude docs

Data Leakage & PIIPreventing exfiltration and protecting sensitive data. OWASP GenAI

Auth & Secrets ManagementScoped credentials, OAuth, and secret handling for tools. MCP docs

Action AuthorizationApproval policies and confirmation for high-risk actions. Anthropic Engineering

Red-Teaming AgentsrecAdversarial testing of tool misuse and jailbreaks. Microsoft docs

OWASP LLM Top 10recStandard threat model for LLM and agent applications. OWASP GenAI

Stage 14 / 17 · 6 topics · 0 lessons

Adaptation & Fine-Tuning

Specializing models for agentic tasks when prompting isn't enough.

When to Fine-TunePrompting vs RAG vs fine-tuning decision framework. OpenAI docs

Supervised Fine-TuningrecBuilding datasets and SFT for task specialization. Hugging Face TRL docs

PEFT & LoRArecParameter-efficient tuning and adapters. Hugging Face PEFT docs

Tool-Use Fine-TuningoptionalTraining models to call tools reliably. Hugging Face TRL docs

RL for AgentsoptionalRLHF, DPO, and RL fine-tuning for agentic behavior. Hugging Face RL course

Distillation & Small ModelsoptionalDistilling capable small models for cost and latency. Hugging Face docs

Stage 15 / 17 · 8 topics · 0 lessons

Production Engineering & Deployment

Shipping agents as reliable, scalable services.

Serving AgentsFastAPI endpoints, async workers, and streaming responses. FastAPI docs

Long-Running & Async JobsQueues, workers, and durable execution for slow agent runs. Celery docs

Docker & DeploymentContainerizing agents and deploying to cloud/Kubernetes. Docker docs Lab

Caching & Cost ControlsrecSemantic caching, budgets, and spend guardrails in prod. Claude docs

Scaling & ConcurrencyrecHandling many concurrent sessions and rate-limit pooling. vLLM docs

Reliability & FallbacksTimeouts, retries, circuit breakers, and model fallbacks. Google SRE Book

CI/CD for AgentsrecEval-gated deploys, prompt versioning, and rollbacks. GitHub docs

Monitoring & Incident ResponserecLatency/cost/quality dashboards and on-call for agents. Langfuse docs

Stage 16 / 17 · 5 topics · 0 lessons

Agentic UX & Interfaces

Designing the human-facing surface where agents earn trust.

Streaming & Progress UXShowing reasoning, tool calls, and partial output live. Claude docs

Transparency & CitationsrecSurfacing sources, confidence, and agent reasoning. Claude docs

Control & InterruptibilityrecLetting users pause, edit, and redirect the agent. LangChain docs

Generative UIoptionalAgents rendering interactive components and forms. Vercel AI SDK docs

Feedback CapturerecThumbs, corrections, and feeding signal back into evals. LangSmith docs

Stage 17 / 17 · 7 topics · 0 lessons

Capstone Projects & Career

Proving job-readiness with portfolio-grade agents and interview prep.

Project: RAG Knowledge AgentProduction RAG agent with citations, eval, and observability. LangChain docs

Project: Coding AgentAutonomous code agent that edits files and runs tests. Anthropic Engineering

Project: Multi-Agent WorkflowrecOrchestrated crew automating a real end-to-end business task. LangChain docs

Project: MCP Tool ServerrecCustom MCP server integrating a real third-party system. MCP docs

Agent System DesignInterview-ready design of scalable, safe agent architectures. Anthropic

Portfolio & WritingrecDemos, write-ups, and open-source contributions that get hired. Hugging Face docs

Staying CurrentrecTracking papers, releases, and the fast-moving agent ecosystem. Hugging Face Papers

You're job-ready.

Clear every stage, earn the certificate, and walk into interviews prepared. The complete path, nothing hidden, no gaps.

Destination reached