Module IV — AI Agents and Orchestration (2 ECTS)

Agents capable of planning, reasoning, and using tools. Multi-agent systems with LangGraph and MCP. Memory, flow control, and reliability evaluation. ~55 hours of work.

What you will be able to do upon completion#

1. Implement the ReAct loop by hand, without frameworks, and explain exactly what each component does.
2. Model agents as state graphs in LangGraph: cycles, conditionals, checkpoints, and human-in-the-loop.
3. Develop and publish a MCP server of your own and connect it to Claude Desktop / Claude Code.
4. Design robust tools: clear contracts, error handling, idempotency.
5. Build a multi-agent system with a supervisor and reason about when (not) it is worth it.
6. Equip an agent with working, episodic, and semantic memory.
7. Explain the architecture of Amazon Bedrock Agents (action groups, KBs) in the context of AIF-C01.
8. Evaluate agents: trajectories, determinism, cost per task, and safety guardrails.
9. Situate RLHF, RLAIF, and Constitutional AI within the alignment and fine-tuning landscape.

Prerequisites#

• Modules I–III completed (specifically: function calling from Module II and embeddings from Module III).
• Group agents dependencies installed from the repo root:

uv sync --extra agents

• .env at the root with OPENAI_API_KEY (the labs use gpt-5.6-luna, verified in August 2026; almost all also accept ANTHROPIC_API_KEY + claude-haiku-4-5). LANGSMITH_API_KEY is optional but recommended from lab 02 onwards to view traces.

Study order and estimated times#

#	Theory	Associated Lab	Hours
1	teoria/01-patrones-de-agentes.md — ReAct, plan-execute, reflection, self-critique	labs/01_react_desde_cero.py	7
2	teoria/02-langgraph.md — state graphs, cycles, conditionals, checkpoints	labs/02_langgraph_basico.py, labs/03_langgraph_ciclos_condicionales.py, labs/04_langgraph_checkpoints.py	10
3	teoria/03-mcp.md — MCP architecture and server development	labs/05_mcp_server.py	6
4	teoria/04-tool-use-avanzado.md — tool design, errors, idempotency	(cross-cutting: applied in all labs)	4
5	teoria/05-multi-agente.md — coordination, supervisors, communication	labs/06_multiagente_supervisor.py	7
6	teoria/06-memoria.md — episodic, semantic, and working memory	labs/07_memoria_agente.py	6
7	teoria/07-bedrock-agents.md — action groups, knowledge bases	no lab (requires AWS account; console flow documented)	3
8	teoria/08-evaluacion-de-agentes.md — determinism, costs, guardrails	(mini-project evaluated with this rubric)	5
9	teoria/09-alignment-y-fine-tuning.md — RLHF, RLAIF, Constitutional AI	no lab (theoretical)	4
—	ejercicios.md — 12 exercises + mini-project (research agent)	Tests and rubric	13

Total: ~55 h (≈ 2 ECTS). At 10 h/week, this corresponds to weeks 13–18 of the program.

How to work through the module#

1. Read the theory from the before block of the lab; each lab assumes the concepts from its theory file.
2. Run each lab, read it entirely, and modify it (each docstring suggests variations). The labs offer offline workflows and have iteration limits. Before using live modes, calculate the cost using the current rate and set a budget.
3. Complete the exercises from the block upon finishing it, not all of them at the end.
4. The mini-project (exercise 12) is the module's milestone: a research agent with 3+ tools, traces, and evaluation. Save it: you will reuse it in the capstone.

Module Mind Map#

flowchart TD
    A[Patrones: ReAct, plan-execute,\nreflection, self-critique] --> B[LangGraph:\nstate, cycles, checkpoints]
    B --> C[Multi-agente:\nsupervisor and communication]
    B --> D[Memoria:\nworking, episodic, semantic]
    T[Tool use avanzado] --> A
    T --> M[MCP: herramientas\nestandarizadas entre apps]
    C --> E[Evaluation:\ntrajectories, cost, guardrails]
    D --> E
    BR[Bedrock Agents\nmanaged version on AWS] -.cloud equivalent.-> B
    AL[Alignment: RLHF, RLAIF,\nConstitutional AI] -.why models\nfollow tool calls.-> A

Honest Warning Before Starting#

Most problems you will encounter in production do not require an agent: a fixed pipeline of 2-3 LLM calls is cheaper, faster, and easier to debug. This module repeatedly emphasizes the question "Do I need an agent here?" — knowing how to answer it is as important as knowing how to build them. Read it with this critical mindset: it is what distinguishes an engineer from someone who just chains demos.