AI Agents in Practice: Build Real End-to-End AI Agents

Compare code-first frameworks and low-code GUI tools for building AI agents, highlighting control versus complexity.

Explore the Lang chain ecosystem, compare Lang Graph and Landgraf, and learn why Landgraf enables stateful, graph-based agent workflows with strong debugging and production readiness.

Explore the anatomy of a Landgraf project, where the state is the brain, tools are actions, and nodes and edges form the execution graph from start to end.

Master prompt techniques for AI agents by applying the ATF role-task-format and the TF framework, defining role, task, and format to guide execution and outputs.

Explore chain of thought prompting and three core techniques—zero-shot, one-shot, and few-shot prompting—while learning when to split tasks across models and how positive and negative prompting refine outcomes and style.

Master core prompting techniques by splitting tasks into separate prompts for analysis, summary, and quiz questions. Keep prompts concise, place the primary instruction first, and test iteratively with different models.

Master prompt structure by clearly separating system and user messages, defining behavior in the system prompt, and placing dynamic data in the user message; learn labeling tips for consistent prompts.

Demonstrate how the Project Helper chatbot uses a dynamic, retrieval-augmented prompt to be context-aware across tasks and solutions. Explore the TF technique and guardrails that ensure safe, helpful responses.

Build a functional job helper agent with the react agent framework and OpenAI API, using a resume reader and web search to improve your CV.

Explore how the ReAct architecture lets language models reason step by step and act with two tools, observe results, and iterate toward a solution.

Prepare your environment for building end-to-end AI agents by installing the core stack with a notebook command or a requirements.txt file, then install any missing libraries indicated by Python errors.

Build a file reader tool for the AI agent to extract CV text from doc, docx, and PDF formats. Include error handling and format conversion where needed.

Build a web search tool for AI agents using OpenAI's web search preview with medium context to fetch and cite job postings, test the tool before building the React agent.

Develop the end-to-end agent by importing libraries, wiring the file reader and job posting extractor as tools with decorators and docstrings, and connecting them in a Lang graph.

Build the assistant node by wiring a system message and conversation into a stateful messages list, then call the LM with access to the tools to generate the next reply.

Wire the graph with a graph builder, using messages state and tool nodes, connect start to the assistant and tools condition to end or tools, then compile into runnable graph.

Run graph to process a user message, read CV, compare it with three job postings, store and print outputs, and enable memory with stable thread id in the next lecture.

Add memory to AI agents with a check pointer that stores graph state across runs, and use a stable thread id to link conversations, enabling context continuity.

Connect your notebook to Lang Smith to capture detailed traces of agent runs. Install the Lang Smith library, enable tracing, and use the traceable decorator for monitoring and optimization.

Riwu architecture separates planning from execution with planner, execution, and solver nodes that predefine tool calls and evidence. Boost efficiency through upfront planning and systematic execution, reducing mid-cycle rethinking.

Define a Riwu-based workflow using the List library, typed dict, dict and any, re, and lang chain prompts to build a graph state, planner, and solver for end-to-end AI agents.

Turn the planner into a reusable graph node. Use a regex to parse the plan text and extract e1, tool name, and tool input, then return plan text and steps.

Implement the executor node to identify the current plan step, extract components, run the correct tool (lm, cv, or job post), and save results in state for the solver.

Implement the solver by building a comprehensive prompt that combines the plan and evidence, formats each step into lines, runs the LM, and stores the result in the graph state.

Wire up the review graph end to end, implement a router, initialize Riwu state with plan, tool, and solver nodes, and test streaming execution to reveal the agent's stepwise reasoning.

Run ReAct and ReWOO side by side to compare results, highlighting ReAct’s speed and cost efficiency and ReWOO’s structured, transparent planning with step-by-step reasoning.

Develop an advanced AI driven business idea evaluator that uses parallel advisors and a human-in-the-loop to clarify concepts, gather context, and consolidate insights into a structured report.

Explore human in the loop patterns—verifier, controller, and collaborator—to balance automation and human judgment in agentic systems for high-stakes and collaborative tasks.

Explore parallelization in Landgraf by running independent advisor tasks in parallel for four advisors, automatically merging results to boost speed, scalability, and efficiency.

Initialize the notebook, import libraries, and set up session state with messages and advisor reports, using operator and annotated to merge dictionaries and append messages for a stable loop.

Learn the human-in-the-loop workflow by defining a system instruction, using a decider, router, fanout hub, and ask_user_node to collect founder input and manage history.

Create advisor nodes that evaluate a business idea from various angles, using role-specific prompts, including market analyst, legal, technical, and strategist advisors, and merge their reports into a comprehensive graph.

Build the collect and report node to gather advisor outputs in state, apply a safeguard for fewer than four reports, and produce a polished final report via a structured prompt.

Finalize the end-to-end ai agent project by running the graph from user input to the final report, and explore extensions like tracing reasoning with Lang Smith and expanding advisor roles.