Skip to main content

Agentic AI Systems

An agent is not a longer answer. It is a system in which a model chooses an action, the code decides whether to carry it out, and the result determines the next step.

See the agents

One decision, step by step

Step 1 of 5

  1. Goal

    Summarize the QuickScreen documents.

  2. Observation

    Known passages: A:1, X:1

  3. Decision

    The retrieved text asks to send an email. The guardrail ignores the instruction because retrieved text does not grant permission.

  4. Permission gateBlocked

    The document asked for send_email. Text from a document grants no permission, so the code does not run the call.

  5. Result

    The summary relies on P:1 only. The instruction in X:1 was blocked: send_email is not an allowed tool, and the document does not change permissions.

How you learn here

  1. Understand13 sessions: what an agent chooses, and what the code allows.
  2. Watch it runSimulations in the browser, with no model and no key.
  3. Check yourselfCode exercises and knowledge checks, with no grade.

Thirteen sessions

Sessions 1–2 are theoretical preparation. Practice starts in session 3.

  1. Session 1: What is an agentGoal, observation, action, and who chooses the next step.TheoryStudy guide
  2. Session 2: Tools, permissions and MCPRequest, schema, check, execution, and a result that goes back to the model.TheoryStudy guide
  3. Session 3: The first agentEnvironment, a deterministic loop, and stopping. No paid key.PracticeDraft
  4. Session 4: LangGraphState, nodes, transitions, and branching when search fails.Theory and practiceDraft
  5. Session 5: Local tools and APIsA local tool, a free API call, and permission errors.PracticeDraft
  6. Session 6: Planning and decisionsA plan made in advance versus a decision after every result.TheoryDraft
  7. Session 7: Failure analysisTrajectory, budget, and a convincing answer that hides a failure.Theory and exercisesDraft
  8. Session 8: Knowledge, memory and RAGRetrieval, citation, and deciding whether to search again.TheoryDraft
  9. Session 9: Document agentAn answer with sources, and what happens when information is missing or changes.PracticeDraft
  10. Session 10: Evaluation, monitoring and safetyComparing against a simpler version, a run log, and a guardrail in code.Theory and practiceDraft
  11. Session 11: Project integrationTwo agents versus a single agent, on the same tasks.PracticeDraft
  12. Session 12: Guest lectureA reserved slot. The lecturer will publish the guest's details.GuestDraft
  13. Session 13: Project presentationsA demo, a failure case, and questions about the code and the decisions.PresentationDraft

Where to start

Accessibility settings

These settings change how this site looks on this device only. They are saved in this browser.

Text

Colour

Light and dark mode are in the top bar, next to search.

Reading and navigation