Agents in production: my class at UNICEN

Originally posted on X.

I gave a class on AI agents at UNICEN, for the systems design course.

The class is in Spanish. The chapters and sources below are in English.

The class goes from a single LLM call to multi-agent systems in production. I tried to explain concepts that I believe are must-haves in today’s work landscape, such as observability and evals, for example. At the end I go through a real case, how we built Lumi at Tiendanube.

Chapters

• 00:00 Intro and what the class covers
• 07:32 Workflows
• 09:13 When to use LLMs, and when not to
• 11:17 Prompt chaining
• 14:31 Routing
• 17:39 Parallelization
• 20:20 Evaluator-optimizer
• 21:34 Limits of workflows
• 22:42 What is an agent? Anatomy, tools and memory
• 25:44 Tools and MCP (Model Context Protocol)
• 34:45 Limits of agents
• 37:32 Observability
• 41:33 Evals
• 46:40 Multi-agent architectures
• 47:32 Subagents
• 49:47 Handoffs
• 50:13 Skills
• 50:52 Router
• 51:32 When to use each architecture
• 52:11 Guardrails and middlewares
• 55:49 Real case

Sources

• Building Effective Agents
• Jev
• Lost in the Middle: How Language Models Use Long Contexts
• Model Context Protocol
• Introducing the Model Context Protocol
• AI Engineer
• A2A: a new era of agent interoperability
• Choosing the right multi-agent architecture
• How we use LangChain to power Lumi for 180k+ merchants
• Scaling LangGraph’s Postgres Checkpointer in Production

Four takeaways

1. Decide first whether the problem needs an agent at all. If it can be solved deterministically, solve it deterministically.
2. Start with the simplest architecture that works, and add complexity only when the problem asks for it.
3. Evaluation, observability and monitoring are the pillars. They are what separates a good agentic system from an average one.
4. Build it so it can scale to real users in production, in a controlled way.