Skip to content
Idaho AI Classroom
All lessons
AI Technical AptitudeAI Methods9-12.AITA.AIM.2

Drop In an Agent

9–12 CS / engineering50 minutes90 minutes (lab extension)

Standard quoted exactly

Integrate a prebuilt AI agent into an application.

Student-friendly learning targets

  • I can specify a prebuilt agent with a job, tools, and a human approval step, then show where it sits in an app.
  • I can integrate on a flowchart (and optional no-code) rather than training an agent from scratch.
  • I can refuse tool actions that send email, change records, or share PII without a human.

Essential questions

  1. What makes something an agent rather than a single model call?
  2. Which tools should an agent be allowed to use without asking?
  3. Where does integration end and 'the agent is the whole product' begin?

Objectives

  1. Spec a help-desk or lab-assistant agent: goal, allowed tools, forbidden tools, escalation.
  2. Draw the host application flowchart with the agent as one box, not the whole diagram.
  3. Require a human approval step before any irreversible action.
  4. Optional: drop the spec into a district-approved no-code agent builder; paper is full credit.

Key vocabulary

Agent
Software that can plan a short sequence of actions toward a goal, often by calling tools (search, lookup, ticket). Prebuilt means we did not train it today.
Tool
A function the agent may call: look up a river gauge, search a FAQ, file a ticket. Tools have side effects; permissions matter.
Integration
Connecting the agent to a host app with inputs, outputs, logging, and approval hooks — not building the agent from zero.
Human approval step
A required pause where a person must click yes before email, record changes, purchases, or alerts go out.
Escalation
Handing the session to a human when confidence is low, the user is distressed, or the tool is out of scope.

Teacher background

AIM.1 was one supervised prediction. AIM.2 is a prebuilt agent that can use tools. Students specify a help-desk (Idaho Power-style outage FAQ) or a CS lab assistant (how to use the laser cutter, where the jumper wires are) and integrate it into an application flowchart. District no-code is optional. Human approval is required by this lesson even if a vendor would skip it. Tools should be boring and public: FAQ search, public USGS lookup, 'create draft ticket.' Forbidden: send email, DM a student, charge a card, unlock a door, scrape a gradebook. Offline: card-sort tools onto allowed / approval / forbidden, then tape them onto the flowchart.

Materials and prep

Materials

  • Agent spec sheet: name, user, goal, 4 tools, 3 forbidden, approval gates, escalation.
  • Flowchart poster of a host app (school work-order site or county extension chat widget).
  • Tool cards: FAQ search, USGS cfs lookup, draft ticket, send email, open locker, read SIS grades, web search (filtered), web search (unfiltered).
  • Optional no-code agent environment if district-approved and account-free for students.
  • Printed AIM.7 preview: safeguards will go deeper later; today the gate is the approval step.

Before class

  • Choose the default story: CS lab assistant for your room, or a public-lands 'campfire rules' help desk. Local beats generic.
  • If no-code is flaky, do not mention it until extension time.
  • Pre-sort one worked example so the approval diamond is obvious on the poster.

Instructional sequence

Model call vs. agent

5 min
  1. AIM.1 recap: one input, one label. Ask what extra happens if the software can search a FAQ, then file a ticket, then stop.
  2. Name that extra: tools + a goal = agent-shaped. We will integrate a prebuilt one, not train it.

Spec, slot, approve

10 min
  1. Three moves: spec the agent (job and tools), slot it into a host app, require approval on side-effect tools.
  2. Demo a bad integration: the agent is the entire school website and can email parents. Demo a good one: chat widget with FAQ + draft ticket, counselor clicks send.
  3. Walk tool cards onto a permission board: auto / ask a human / never.
  4. Say the sentence students must write: 'No send without a human.'

Lab-assistant spec as a class

12 min
  1. Fill the spec sheet together for a lab assistant: goal = help students find procedures and report broken equipment.
  2. Sort eight tool cards. Debate web search filtered vs. unfiltered (AIM.7's child-appropriate example, previewed).
  3. Draw the host app: student opens a help tab → agent → draft → teacher approval → ticket system.

Your help desk

10 min
  1. Pairs spec a second agent: county extension crop FAQ or outage help desk. Complete tools, forbidden list, and two approval diamonds.
  2. They draw the host application around the agent (map, phone tree, kiosk).
  3. Optional no-code: paste the spec into the allowed builder; still print the flowchart.

Real-world examples

  • A hospital scheduling agent that drafts appointments but cannot confirm without a clerk — same pattern as our ticket gate.
  • Idaho Power outage chat that can look up a public map and open a ticket, with a dispatcher approving crew messages.
  • A school library agent that searches the catalog (tool) but cannot charge a late fee.
  • Extension 'ask an expert' widgets that retrieve public fact sheets and queue a human agronomist.

Hands-on activity

Tape the agent into the app

8 min
  1. Pairs tape tool cards onto the host flowchart with auto/ask/never stickers.
  2. A neighbor red-teams: find a path that emails without approval. Fix it.
  3. Initial the 'No send without a human' line.

Discussion questions

  1. If the agent can search the web, is that a tool we auto-allow in a 9th-grade lab?
  2. Does integration mean the agent should remember students across weeks? (Privacy no.)
  3. What is the smallest host app that still counts (a paper kiosk form + agent box)?
  4. When should the agent shut up and escalate?

Differentiation

Support

  • Pre-written goal sentence; students only sort tools and place two diamonds.
  • One host flowchart template with blank boxes.

Challenge

  • Add a tool that looks up public USGS cfs and specify the cache/timeout behavior.
  • Write a one-paragraph 'agent contract' a principal could sign.

Multilingual learners

  • Tool names with glosses; spec goal may be bilingual; forbidden list in both languages.
  • Role-play the approval step in the student's preferred language.

IEP / 504

  • Card sort is the assessment path; drawing can be a partner's job.
  • No live chatbots if they are overstimulating; paper agent only.

Assessment

Formative

  • Tool sort on auto/ask/never.
  • Class flowchart with a visible approval diamond.

Summative

  • Spec sheet + host flowchart with the agent integrated, two approval gates, and a forbidden-tool list.
  • A chatbot doodle with no tools and no host app does not meet 'integrate into an application.'

Success criteria

  • Agent is prebuilt/specified, not trained in class.
  • At least one tool and one forbidden action are named.
  • A human approval step sits in front of side effects.

Responsible use, ethics, and privacy

Responsible use

Prebuilt, district-approved, or paper. No student-built agents with unsupervised web tools. No connecting to school email.

Ethics

An agent that 'just helps' can still harass, over-promise, or skip a human. Integration includes saying no to tools.

Privacy

FERPA: the agent may not read SIS, IEPs, or student mail. Sessions should not store names. Public FAQs only.

Reflection

  1. Where in your flowchart is the agent, and where is the application?
  2. Which tool was the hardest to classify as auto/ask/never?
  3. What would break if we removed the approval diamond 'just for speed'?

Homework

Finish the spec sheet for a campfire-rules help desk on public-land FAQs. List four tools, three never-tools, and the exact sentence the human sees on the approval screen. Paper only.

Closing

You did not train an agent. You specified one, gave it tools, and bolted it into an app behind a human gate. That is integration. Next we look at how a model writes a function from data at all.

Extensions and cross-curricular links

Go further

  • 90-minute block: tabletop-test the flowchart with another pair playing user, agent, and approver; log failures.
  • No-code drop-in if district-approved; still keep the paper contract.
  • Python extension: a stub function approve(action) that returns False unless a teacher flag file exists — conceptual, not production.
  • Preview AIM.7 safeguards without stealing that lesson's planning depth.
Workplace skills
Help-desk runbooks already have escalation; agents must inherit them.
Civics
A county chat agent that can file reports is a public service and needs public oversight.
Theater / ELA
Tabletop role-play of user/agent/approver makes the integration visible.