Agentic AI
Where Generative AI builds the model layer, Agentic AI builds autonomous systems on top of it: the patterns (reflection, tool use, planning, multi-agent), the protocols (MCP, A2A), the orchestration frameworks and SDKs, and the production concerns for agents that actually work.
- Chapter 01
Foundations
18 lessons - 01 What agentic AI means An agent is an LLM that can act, not just answer. Here's what that actually involves — and how it changes the engineering.
- 02 Agentic design patterns Anthropic's canonical taxonomy — prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer — and the agent shape itself.…
- 03 ReAct, Plan-Execute, Reflexion The three agent reasoning loops every engineer should know in 2026 — interleaved ReAct, upfront Plan-and-Execute, and self-correcting Reflexion. Ho…
- 04 ReWOO: plan-execute without observation ReAct re-reads the whole growing transcript on every step — expensive and easily derailed. ReWOO plans the entire task up front, runs the tools, th…
- 05 HTN & evolutionary planning Free-form LLM planning is flexible but unpredictable. Hierarchical Task Network planning brings classical-AI structure — decompose a goal into subt…
- 06 Reflection Anthropic's name for the reflection pattern. How to make an LLM check its own work, the two ways to wire it, when same-model is enough vs. when a s…
- 07 Tree of Thoughts Chain-of-thought reasons in one straight line — if an early step is wrong, the whole chain fails. Tree of Thoughts turns reasoning into deliberate…
- 08 Skill libraries & sleep-time compute Most agents re-derive the same procedures on every task. A skill library lets an agent solve something once, save the solution as a reusable skill,…
- 09 Self-improving agents Most agents are static — fixed weights, fixed skills. A self-improving agent closes a loop: generate attempts, evaluate them, keep and learn from t…
- 10 Tools — defining, calling, validating Defining tools with JSON schema, validating arguments, handling errors, and the production patterns for reliable tool-using agents.
- 11 AGENTS.md, Skills & Tools Three different ways to give an agent knowledge and power — always-on instructions, on-demand model-loaded skills, and runtime tool calls. What eac…
- 12 The Agent Harness The LLM is stateless — the harness is everything else. Learn what the runtime scaffolding around a model is, what each component does, and how to b…
- 13 Model Context Protocol (MCP) The emerging standard for plugging tools into agents. Write your integration once; every MCP-compatible client can use it.
- 14 Advanced MCP primitives Tools, resources, and prompts are only half of MCP. The protocol is bidirectional and interactive: a server can call back to the host's LLM (sampli…
- 15 MCP gateways & registries One MCP server is easy; an organization with dozens of servers and many clients is a mess. An MCP gateway is the fix — a proxy that aggregates serv…
- 16 MCP vs A2A vs ACP vs ANP Four agent protocols arrived in barely a year — and most takes get them wrong. They aren't competitors; they're layers. Here's how MCP, A2A, ACP, a…
- 17 A2A — Agent2Agent Protocol MCP connects one agent to its tools. A2A connects independent, opaque agents to each other. A deep dive on the Agent Card, the task lifecycle, stre…
- 18 How agents authenticate each other When agents call other agents and tools on your behalf, 'it said it was the billing agent' is not security. Identity, authentication, scoped author…
- Chapter 02
Multi-agent systems
6 lessons - 19 Multi-agent: supervisor & swarm Supervisor, swarm, and hierarchical orchestration — and the harder question of when NOT to go multi-agent. Most tasks are better served by a single…
- 20 Multi-agent debate & voting A single agent can be confidently wrong. Ask several and either let them vote or let them argue: independent errors cancel under majority vote, and…
- 21 Group chat & the blackboard pattern Agents on a team have to do two things: share what they know, and decide who acts next. There are two classic ways to wire that — broadcast group c…
- 22 Agent negotiation & economies Orchestration assumes one owner and a shared goal. But agents increasingly meet across that boundary — your buyer agent and someone else's seller a…
- 23 Generative agents: simulations Most agents do tasks. Generative agents do something stranger — they simulate believable human behavior. Park et al.'s Smallville put 25 LLM agents…
- 24 Why multi-agent systems fail Multi-agent systems fail more often than their hype suggests — and not just because the models are weak. Coordination itself adds new failure modes…
- Chapter 03
Multi-agent theory
4 lessons - 25 Consensus & Byzantine fault tolerance The LLM-agent lessons assumed cooperative agents under one owner. Classical multi-agent systems studied a harder problem: how do distributed, possi…
- 26 Multi-agent reinforcement learning Single-agent RL has one learner in a fixed environment. Multi-agent RL has many agents learning at once — which breaks that assumption. From any on…
- 27 Swarm intelligence & optimization Most multi-agent design has agents pursuing goals. Swarm intelligence is a different angle: global optimization that emerges from many simple agent…
- 28 Theory of mind for agents To coordinate, compete, or communicate well, an agent benefits from modeling what other agents think — their beliefs, goals, and knowledge, which m…
- Chapter 04
Orchestration frameworks (LangChain, LangGraph, LlamaIndex)
11 lessons - 29 Models, prompts, parsers The minimum-viable LangChain — chat models behind a unified interface, prompt templates, output parsers, and when you actually need any of it.
- 30 Chains & LCEL LangChain Expression Language is the pipe-operator DSL for stitching Runnables together. Branching, fan-out, streaming, batching — all on the same…
- 31 RAG with LangChain LangChain's RAG pieces fit together cleanly for prototypes. Document loaders, text splitters, embeddings, vector stores, retrievers — wired end-to-…
- 32 Build agents from scratch Explicit state graphs for the workflows LCEL can't express cleanly — branching, loops, retries, and pauses. The mental model and the smallest possi…
- 33 Nodes, edges, state State is a TypedDict; nodes return partial updates; reducers decide how to merge. The `add_messages` reducer is the building block for every chat-s…
- 34 Persistence & streaming Checkpointers save state after every step — enabling resumable agents, multi-turn conversations, time-travel debugging, and streaming intermediate…
- 35 Human-in-the-loop Pause the graph, let a human review or edit, resume from where it stopped. The pattern is "LLM proposes, human disposes" — and it's what makes agen…
- 36 Indexes, query engines & retrievers LlamaIndex is the data framework for LLM apps — it turns documents into a queryable index through swappable components: loaders, node parsers, inde…
- 37 Event-driven Workflows Workflows are how modern LlamaIndex composes multi-step LLM apps: typed Events flowing between async Steps. The recommended way to build everything…
- 38 FunctionAgent & ReActAgent Turn a LlamaIndex index into a tool an agent can call. FunctionAgent (native function-calling) vs ReActAgent (text-based ReAct), both built on Work…
- 39 LlamaParse — document parsing Naive PDF text extraction destroys tables, columns, and layout — and poisons your RAG. LlamaParse uses vision models to turn messy documents into c…
- Chapter 05
Agent SDKs & runtimes
13 lessons - 40 Agents, handoffs & guardrails A lightweight, production-minded agent framework: Agents, Runner, Tools, Handoffs, Guardrails, and Sessions. The default starting point for many pr…
- 41 MAF overview & first agent MAF is Microsoft's unified agent framework, replacing AutoGen and Semantic Kernel. Here's the shape of a first agent and the four building blocks y…
- 42 Adding tools How Python functions become MAF tools, when to require human approval, and the hosted tools (code interpreter, file search, MCP) you get for free.
- 43 Workflows MAF workflows are graphs of agents (executors) for sequential, parallel, conditional, and looping orchestration. Here's a 3-agent planner → executo…
- 44 Middleware & guardrails Middleware in MAF wraps around agent invocations and tool calls. Use it for logging, rate limiting, PII scrubbing, and termination. Order matters.
- 45 Install & first agent Google's Agent Development Kit is the Gemini-native agent framework. LlmAgent, tools, sessions, runners — plus a web UI for local dev. Here's the s…
- 46 Multi-tool agent Defining tools in Google ADK, using multiple tools in one agent, and the built-in tools (Google Search grounding, code execution) you get for free.
- 47 Workflow patterns ADK gives you SequentialAgent, ParallelAgent, and LoopAgent for multi-agent orchestration. Here's a research → write → review pipeline.
- 48 Deployment Three paths from local ADK agent to production. Agent Engine for fastest, Cloud Run for portability, GKE for full control. The trade-offs, with com…
- 49 CrewAI — role-based crews CrewAI models a multi-agent system as a team you hire: each agent gets a role, a goal, and a backstory, each does a task, and a Process decides how…
- 50 Claude Agent SDK The Claude Agent SDK is the production harness behind Claude Code, available standalone. It ships the loop that already works — gather context, tak…
- 51 AutoGen: conversational multi-agent Where LangGraph models agents as a graph and CrewAI as roles, AutoGen models them as agents that talk. Microsoft's framework for multi-agent conver…
- 52 Agent runtimes: Agno & Mastra The first agent frameworks were orchestration libraries you wire together. A newer class — agent runtimes — ships the whole production stack (model…
- Chapter 06
Agents in Production
13 lessons - 53 Agent Memory How agents remember across turns and sessions. The cognitive-science taxonomy production systems borrow — working, episodic, semantic, procedural —…
- 54 Context engineering The headline 2026 agent skill: curating what's in the context window. Compaction, isolation, and retrieval keep long-running agents lean, coherent,…
- 55 Agent Security An LLM can't reliably tell instructions from data, so any content it reads can hijack it. Learn the lethal trifecta, OWASP excessive agency, and wh…
- 56 Token theft & runtime authorization Keep API keys and OAuth tokens out of model context, then authorize every tool call at runtime. Build a safe fake-token lab and learn vault injecti…
- 57 MCP tool poisoning & supply-chain security Threat-model the full MCP trust boundary: malicious tool metadata, cross-server shadowing, post-approval rug pulls, confused-deputy authorization,…
- 58 Evaluating agents An agent has a loop, side effects, and non-deterministic intermediate states — so a single input/output check isn't enough. Evaluate the trajectory…
- 59 Agent benchmarks & eval-driven dev Static QA benchmarks don't measure an agent — you need end-to-end task completion in a real environment. SWE-bench, GAIA, WebArena, OSWorld, why ag…
- 60 Observability & tracing You can't debug an agent you can't see. Traces capture every LLM call, tool call, token, and millisecond of a run — so you can find the retry loop,…
- 61 OpenTelemetry for GenAI Agents are hard to debug — non-deterministic, multi-step, and spread across model calls, tools, and retrieval. OpenTelemetry's GenAI semantic conve…
- 62 Cost & latency control Agents make many model calls per task, so cost and latency multiply fast. The levers that cut agent spend 40–70%: routing, caching, token budgets,…
- 63 Durable execution for agents An agent run isn't a request-response — it's a long workflow of LLM calls, tool calls, waits, and approvals. If the process crashes mid-run you los…
- 64 Agent safety controls Prompt-level safety asks the model to behave — not enough for an autonomous agent that can delete data, send money, or deploy code. You need engine…
- 65 The coding-agent workbench A chatbot answers; a coding agent must ship working code to a real repository — and that's where reliability collapses. The fix isn't a smarter mod…
- Chapter 07
Frameworks & Tooling
6 lessons - 66 Mem0 — a memory layer for agents How Mem0 extracts salient facts from conversations and retrieves only the relevant ones at the next turn, giving agents durable personalized memory…
- 67 FastMCP — build MCP servers fast A Pythonic framework that turns ordinary functions into MCP tools, resources, and prompts with simple decorators — so you can expose your own data…
- 68 Code execution with MCP Instead of loading every tool and routing every result through the context window, let the agent write code that calls tools out-of-context. The pa…
- 69 Firecrawl — web data for LLMs & RAG Turn any URL into clean, LLM-ready Markdown. Firecrawl handles JS rendering, boilerplate removal, pagination, and full-site crawls so your RAG pipe…
- 70 Computer-use & browser agents Most software has no API. Computer-use agents drive a screen the way a person does — read a screenshot, click, type, repeat — so they can operate a…
- 71 Voice agents (realtime) A text agent has seconds to respond; a voice agent has about half a second, or the conversation feels broken. Real-time voice is its own discipline…
- End of section 0 / 71 complete
Make it stick — pass every quiz.
Each lesson has a short quiz at the bottom. Passing the quiz is what marks the lesson complete and counts toward your certificate.
Agentic AI — frequently asked questions
Straight answers to the questions people ask most about agentic ai.
What makes an AI agent different from a chatbot?
An agent doesn't just answer — it plans and acts in a loop: it calls tools, observes the results, and decides the next step until a goal is met. A chatbot produces one reply; an agent can search, run code, query a database, and chain those steps autonomously.
What is the Model Context Protocol (MCP)?
MCP is an open standard for connecting AI models to tools and data sources through a uniform interface, so any MCP-compatible client can use any MCP server. It acts like a universal adapter that replaces bespoke, per-integration glue code for agent tooling.
What's the difference between LangChain and LangGraph?
LangChain offers building blocks and chains for composing LLM apps; LangGraph models an agent as an explicit graph of nodes and edges with shared state, which makes loops, branching, retries, and human-in-the-loop control far easier to reason about. Use LangGraph when the control flow gets complex.
Why do multi-agent systems often fail?
Common failure modes are agents losing shared context, compounding each other's errors over long chains, looping without progress, and ballooning cost and latency. Reliability usually comes from constraining each agent's scope, adding verification steps, and keeping the control flow explicit rather than fully open-ended.
What are the main agent design patterns?
The core patterns are reflection (the model critiques and revises its own output), tool use (calling external functions), planning (decomposing a goal into steps), and multi-agent collaboration (specialised agents working together). Most production agents combine a few of these.