Agentic Design Patterns: From the Loop to Production

Design, build, evaluate, and ship AI agents with the patterns real teams use: ReAct, planning, reflection, routing, supervisors, guardrails, memory, and evals.

Most people learn agents backwards: they pick a framework, copy a demo, and end up with a system they cannot explain or debug. This course goes the other way. You start with the plain agent loop, learn when a workflow beats an agent, then build the five core patterns one by one on the same project. You add the control layer (guardrails, human approval, budgets), memory and agentic RAG, an eval suite with traces, and a production runtime. You finish with a support agent you can defend in a design review, and every lesson comes with a visual pattern card to keep.

What you will be able to do

• Choose the simplest pattern that solves a problem, and defend the choice
• Build the five core agent patterns from a plain loop, without a framework
• Add guardrails, human approval, and budgets enforced in code
• Evaluate trajectories with a calibrated judge and ship behind traces

Before you start

• Comfortable with Python and calling an API
• You have built a small LLM app (RAG or a chatbot)

Course outline

1. Workflow or agent? The autonomy ladder: The same task three ways: one call, a chain with gates, and a small agent, measured side by side. A decision checklist you can use in a design review.
2. The five core patterns: ReAct with typed tools, plan-and-execute, reflection with a test-based critic, routing, and supervisor-workers on one support project. A pattern starter repo, written without a framework.
3. More patterns: workflows and reasoning: Prompt chaining, parallel voting, evaluator-optimizer, ReWOO, Reflexion, and code as action in a sandbox. A pattern picker for your own problems.
4. Control the loop: Input, action, and output guardrails, a human approval step with checkpoints, and budgets in code. A security review using the lethal trifecta.
5. Memory and agentic RAG: Episodic and semantic memory with a write policy, graded retrieval, and context compaction. An agent that is useful across sessions, with answers traced to sources.
6. Evaluate and ship: A trajectory eval set, a calibrated LLM judge, traces, and a queue-based runtime with retries and fallbacks. A release report: success rate, pass^k, cost per successful task.
7. Capstone: a production support agent: Routing, specialist agents, agentic RAG, guardrails, human approval for refunds, and an eval suite in CI. A deployed agent and an architecture you can defend in an interview.

All AI and machine learning courses