AI agents: design, build & operate
The vendor-neutral craft of agents: what an agent really is, the design patterns that actually work, how to give it tools, memory and retrieval, how to coordinate several of them, and how to evaluate, guard, and ship them to production.
Everyone is talking about AI agents. Most of them cannot tell you what an agent actually is, let alone ship one that survives contact with production. This block fixes that. You will learn what an agent really is, the perceive, plan, act, observe loop that separates true agents from glorified scripts, and how to decide when you need an agent at all versus a plain workflow or automation. Autonomy is a dial, not a switch, and you will learn to set it deliberately.
We go straight into the patterns that matter: tool loops, ReAct, planning, reflection, and routing. Then we build. You will give your agent hands through tools and function calling, wire up short-term context and long-term recall, run retrieval inside the agent loop with agentic RAG, and coordinate multi-agent systems with orchestrators, workers, and clean handoffs. This is the difference between a demo that impresses a boardroom and a system that quietly does the work.
The part almost nobody teaches is operations, and it is where careers are made or broken. You will evaluate and debug agents using traces, evals, and a real understanding of failure modes. You will set guardrails, permissions, and human-in-the-loop checkpoints so your agents stay on a leash. And you will manage the trio that decides whether an agent ever ships: cost, latency, and reliability.
As a senior leader you do not need to write the code. You need to know what good looks like, where the risk lives, and which questions expose a team that is guessing. This block gives you exactly that leverage. You will walk into any agent conversation as the sharpest person in the room.
What you'll master
- Define what an AI agent is and map the perceive, plan, act, observe loop with confidence
- Choose the right level of autonomy between agents, workflows, and automations for any use case
- Apply core agent design patterns including tool loop, ReAct, planning, reflection, and routing
- Build agents that use tools, retain memory, and run retrieval inside the loop with agentic RAG
- Orchestrate multi-agent systems with clear roles, workers, and handoffs
- Evaluate and debug agents using traces, evals, and structured failure analysis
- Ship agents to production while controlling cost, latency, reliability, and human oversight
Modules
Covers what agents are, when to use them, and the core design patterns behind them.
Covers how to assemble a working agent with tools, memory, retrieval, and multi-agent coordination.
Covers how to evaluate, safeguard, and reliably run agents in production.
Frequently asked questions
What does the AI agents block actually cover?
It covers the full craft of building agents, vendor-neutral: what an agent is (the perceive, plan, act, observe loop), the design patterns that work in practice, how to give an agent tools, memory and retrieval, how to coordinate several agents, and how to evaluate, guard and ship them. It is organised in 3 modules and 10 lessons, from foundations to production operations.
Do I need to be a developer to follow it?
No. The lessons explain the mechanics and the trade-offs of agent systems rather than walking through code in a specific SDK. Someone who has to decide whether to fund an agent project, scope it, or challenge a vendor's claims will get more out of it than from a framework tutorial.
What is the difference between an agent, a workflow and a simple automation?
The difference is how much of the decision is left to the model. An automation follows a fixed sequence, a workflow chains steps with some branching you defined in advance, an agent decides at each turn which tool to call and when to stop. A dedicated lesson, "Agents vs workflows vs automations", covers how to pick the right level of autonomy, because the cheapest reliable option is usually the least agentic one.
Where should I start if my project already works with a single prompt?
Start with the Agent Foundations module before adding anything. It sets out the perceive, plan, act, observe loop and the patterns available (tool loop, ReAct, planning, reflection, routing), which usually shows whether your case needs an agent at all or just a better-structured workflow.
Are multi-agent systems covered, or only single agents?
Both. The "Building an agent" module ends with a lesson on multi-agent systems: the orchestrator and worker split, and how handoffs between agents are organised. It comes after tools, memory and agentic RAG, in that order, because a multi-agent setup built on shaky single-agent foundations multiplies the failures rather than solving them.
What does the production module say about controlling cost and failures?
The third module, "Operating agents in production", devotes three lessons to it: evaluation and debugging through traces, evals and known failure modes; guardrails, permissions and human-in-the-loop; then cost, latency and reliability at shipping time. These are the topics that decide whether a demo that works ever becomes a system you can run.