The Agent Builder is a canvas where students drag blocks and connect them into an AI agent that runs with a real model. This guide covers a first session, the ideas behind the blocks, how to assign a project, and how to read what students turn in.
Read "Build your first agent" first. It is the tutorial your students follow, and it takes about 30 minutes to do yourself.
What students can and cannot do#
Students build from a fixed set of typed blocks. They can pick the model, write the instructions, wire in approved tools, and read every step of a run. They cannot write code that executes, reach arbitrary websites, or use their own credentials. The approved tools without any extra packs are Calculator, Find a Fact, Read File, Write File, Get Date & Time, and Random Number. File tools only reach a folder inside the student's own project.
Every run has limits: two minutes, 20 model calls, 16 tool calls, and a cost cap. A run that hits one stops with a message that says which limit.
Runs draw credits from your class balance. Students never buy credits. If a student sees "You have run out of AI credits", the class balance is empty. Your plan's monthly credits top it back up. On a school plan you can also grant credits to one student from the Students section.
A first session in 45 minutes#
This plan assumes students have accounts and are in your class.
Minutes 0 to 5. Ask: what does a chatbot do between reading your message and answering? Collect guesses on the board. Keep them.
Minutes 5 to 15. Students open Agent Builder, press Create agent, pick One agent that answers, and run it once with any question. Walk the room. The two things that go wrong here are a student who has not pressed Use agent, and a student typing in the wrong box.
Minutes 15 to 30. Students open Base Agent, change the System Prompt, and run again. Give a constraint, not a free choice. "Make it answer only in questions" or "make it refuse to talk about anything but the ocean" work well. Ask two students to read their prompt and their agent's answer aloud.
Minutes 30 to 40. Everyone asks their agent "What is 4,872 times 391?" and checks it with a calculator. Most agents get it wrong. Add the Calculator tool by dragging it onto Base Agent's tray, run again, and open Debug to see the tool call. This is the moment the model stops looking like magic.
Minutes 40 to 45. Go back to the board. Which guesses matched what the Debug tab showed?
Fast finishers: delete a cable, run it, and explain the "Stopped here" dialog to a neighbour. Or try Agent that checks its answer from the starter list.
The ideas behind the blocks#
Students only need the first three of these for a first session. The rest map onto chapter 15, "How AI Agents Work", if you teach it.
- On Message, Base Agent, Show User. A run starts with a message, one block decides what to do, and one block shows the result. Everything else is detail inside Base Agent.
- System Prompt and Call Model. The instructions and the model call. Open Base Agent to see them. Temperature under Advanced is a good first experiment: set it to 0 and the same question gets the same answer.
- Tools. A tool is offered to the model, the model asks for it, the result comes back. Three cables, and the model only knows a tool exists if its Tool Definition reaches the agent's Tools port. Docking makes all three connections at once.
- Limit History. Trims what the model gets to see. The starter "Agent that remembers a little" keeps four turns. Ask students to have a long chat and notice what the agent forgets.
- Compare, If, Template Text, Make Message. Small steps between blocks. The starter "Agent that checks its answer" uses them to route an answer down one of two paths.
- Detect Repetition and Loop Guard. For agents that call the same tool over and over. Advanced.
- Note. A comment on the canvas. It never runs. Ask students to add one explaining their design before they turn in.
Extra blocks come in packs a student imports from the palette. Some packs are part of a paid plan.
Assigning a project#
Projects attach to an activity assignment. Open Assign a lesson, pick a single activity, and set Project submission to one of:
- Saved project. The student turns in the agent as it is when they submit. Good for "build an agent that does X".
- Project run. The student picks one finished run and turns that in with the agent. Good when you want to see a specific conversation, for example "show me a run where it uses the calculator".
The student sees "Turn in an agent project when you are done" on the assignment. They press Turn in or share in the editor, choose the assignment, and press Submit agent.
A submission is a snapshot. Editing the agent afterwards does not change what you see.
Reviewing what students turn in#
Open the assignment from Assignments. The Submitted agents section lists each student with when they submitted. Press one to open the agent read-only. The header says whose agent it is.
Three things worth looking at:
- The canvas. Is a tool docked or wired all three ways?
- See runs in the header, then the Runs tab. Pick a run and read the Debug timeline. It shows every model turn and tool call.
- The System Prompt inside Base Agent. It is where the student's thinking usually shows.
Mark reviewed records that you looked. Write a note sends the student a short piece of feedback and marks it reviewed. The student gets it as a notification. Neither is a grade.
You can also see agents without an assignment when a student shares one with the class. Shared agents open read-only from the link.
What to watch for#
- A student whose agent "does nothing" almost always has an unconnected port. Base Agent shows a dashed dot and "Needs a connection" on it.
- A student who says the calculator is wired but the model never uses it has connected Tool Call and Tool Result but not Tool Definition. Docking avoids this.
- Long, vague system prompts produce loops. "Answer, then stop" is a useful phrase to give them.
- The AI activity section of your class shows every conversation a student has with their agents, read-only, and any moderation flags.
Getting help#
Press Help in the sidebar and send a message. Include the student's username and the name of the agent. We can look at the same run you are looking at.