Language model (LLM)
Language helper: A model that processes or generates language. The simulator uses scripted messages, not a real LLM.
← AI AcademyAI Creators · Grades 7–9
Build, question, test and control
Build a Safe AI Agent.
Help a fictional school library find books about space. See how goals, plans, information, reusable instructions and human permission work together.
Family betaGrades 7–9 with adult supportNo accounts or paid APIsFictional data only
Think first
What turns an answer into an action?
A language model can help interpret a goal. An agent coordinates steps, sometimes using tools. Giving software permission to change something is different from asking it a question.
Learning challenge: Find three suitable books, handle missing information honestly, and refuse to make changes without human permission.
Meet the parts
Five useful ideas, one safe workflow.
These are analogies to support learning, not a claim that AI works like a human body.
Retrieval (RAG-like)
Library lookup: A system can fetch relevant reference material before answering. Retrieval alone does not guarantee truth.
Tool connection (MCP)
Standard connector: MCP is one way to connect AI applications to tools and information. Not every agent uses MCP, and this demo does not use it.
Reusable skill
Recipe: A repeatable set of instructions. Our library skill checks relevance, reading level and the record of each book.
AI agent
Goal-directed helper: Coordinates tasks and tools within permissions. A human can approve, reject, correct or stop risky actions.
The interactive experiment
Mission: Find three books about space.
Choose a situation and step through a pretend agent's plan. Repair failures or control its permission.
Search a fictional catalogue and recommend three books.
PlanSearchCheckApprove if needed
Current step: Ready to start
Select a mission, then start the simulation.
Agent activity log
- No actions yet.
Selected catalogue records
- The helper has not selected books yet.
Why this step matters
No AI model or external library is contacted.
About this simulation: Every title and catalogue ID is invented for this lesson. All decisions are deterministic and happen in the browser. This is not a real LLM, RAG service, MCP connection or autonomous agent. Nothing is reserved, sent or deleted.
Check your understanding
Where should reliable book details come from?
Imagine a helper is asked to recommend a book that does not exist in its catalogue.
Choose what a responsible agent should do.
Continue learning
From understanding to engineering.
In a future Grades 10–12 course, learners could examine actual tool protocols, safety traces, approvals, retries, permissions, evaluations and failure recovery. That technical course is not released.