Understanding AI Agents and Agentic AI, Design, evaluate and secure AI agents from primary sources: workflows, tools, MCP, memory, multi-agent and oversight.
Course Description
Disclaimer: This course contains the use of artificial intelligence
Learn the conditions, not the frameworks. Agent frameworks get rewritten every few months. The design decisions underneath them do not. This course teaches those decisions, taught straight from the primary sources: Anthropic’s and OpenAI’s engineering guidance, the Model Context Protocol specification, OWASP security guidance, and research such as ReAct, Reflexion, SWE-bench and τ-bench.
Every lesson follows the same shape: two defensible designs, and the condition that decides between them. Should this be an agent or a fixed workflow? Should you consolidate tools or minimise them? Retrieve context up front or just in time? One agent or several? You will leave with a test for each, not a slogan.
What the course covers
- What an agent is (and isn’t): the loop, who ends it, and three reasons to say no.
- Workflow patterns: chains, routes, fan-out, orchestrators, evaluators and ReAct.
- Tools and the agent-computer interface: tools as contracts, naming, errors, and code instead of tool calls.
- Model Context Protocol: hosts, clients and servers, who controls what, and what the spec actually requires.
- Context and memory: context as a finite resource, compaction, notes, sub-agents, and memory poisoning.
- Multi-agent systems: what they really buy, manager vs handoffs, and delegation as a trust boundary.
- Evaluating agents: grading final state, pass@k vs pass^k, and building an eval set.
- Security: prompt injection, the trifecta, excessive agency, and containment.
- Humans in the loop and long-running agents: approvals, approval fatigue, stale harnesses and compounding errors.
How you will learn
Short, focused lessons of about four minutes, in-video decision points where you commit to a design before the answer is revealed, live demos, a recap at the end of each module, and six narrated branching scenarios set in a fictional travel company, where each choice changes what breaks next.
The course closes with an agent design review: the conditions from every module, as questions you can ask of any agent design at work. No framework or SDK is required, and nothing here depends on one.
