Skip to content
Idaho AI Classroom
All lessons
AI FluencyEvaluation and Critical Thinking9-12.AIF.ECT.7

What an Agent Actually Does

9–1050 minutes90 minutes (extend hands-on and discussion)

Standard quoted exactly

Recognize and understand how AI agents can accomplish a defined task.

Student-friendly learning targets

  • I can describe an AI agent as a system that pursues a defined task by planning, using tools, checking, and looping, not as a magic person.
  • I can trace a printed agent log and mark goal, plan, tool call, result, and human approval.
  • I can tell a task an agent could attempt from a task that still needs a human on site.

Essential questions

  1. What is the difference between a chatbot that answers and an agent that does steps?
  2. What has to be defined before we let a system act?
  3. Where must a human stay in the loop when an agent books, orders, alerts, or files?

Objectives

  1. Students will define an AI agent using goal, plan, tools, checks, and loops in student language.
  2. Students will label those parts on a fake agent log for an Idaho task.
  3. Students will simulate an agent on paper with tool cards and a human approval gate.
  4. Students will judge whether a task is defined enough to attempt and what failure looks like.
  5. Students will contrast agents with ordinary chatbots and with fully human workflows.

Key vocabulary

AI agent
A system designed to pursue a defined task by planning steps, using tools or apps, checking results, and adjusting, with more action than a single chatbot reply.
Defined task
A job with a clear goal, limits, and a way to tell when it is done, such as 'find three open campsites under $40 on Friday' rather than 'plan a great trip.'
Tool use
When an agent calls another system, such as a calendar, search, form, or inventory list, instead of only generating a paragraph.
Plan-act-check loop
The cycle of making a plan, taking a step, reading the result, and deciding the next step or stopping.
Human-in-the-loop
A required human review or approval before the agent sends, spends, alerts, or files.
Automation
Using technology to carry out steps with less hands-on work. An agent is one form; a dishwasher is another.
Guardrail
A limit on what the agent is allowed to do, such as 'do not submit the permit' or 'do not spend money.'

Teacher background

ECT.7 is recognition and understanding, not a build project. Application standard AP.3 will ask students to construct agents later. Here, ninth- and tenth-graders need a clear picture of how an agent accomplishes a defined task: a goal, a plan, tools, a check, and a loop, with a human still holding the halt sign. A chatbot answers. An agent tries to do. That difference is the whole hour. Keep examples local. A county IT helpdesk agent can reset a password queue. An Idaho Parks agent can look for a campsite. A hospital discharge agent can assemble a checklist. An irrigation agent can draft a pivot schedule. An HVAC shop agent can search a parts catalog. None of them should click Submit, spend money, or page a crew without a person. Students will over-anthropomorphize ('it wanted to help'). Push them back to logs: tool called, result returned, next step. Use a teacher-projected demo if a district agent tool exists; otherwise a printed fake log is better than a flashy video. Paper tool cards let every table simulate the loop on Chromebooks or with no devices at all. End with a defined-task test: if you cannot write the finish line, you are not ready to let a system act.

Materials and prep

Materials

  • Projector and a saved or live teacher demo of a harmless agent-style log (search, then draft, then wait)
  • Printed fake-output packet: agent log for a campsite search and a messy, undefined 'make it nice' task
  • Tool cards: Search calendar, Check fire restriction, Inventory lookup, Draft email, Submit form, Halt for human
  • Role cards: Agent, Tool, Human approver, Auditor
  • Optional Chromebooks; paper simulation is the core
  • Exit tickets

Before class

  • Print logs with obvious steps and one bad action the human must catch (agent about to submit a permit or charge a card).
  • If no district agent tool exists, do not improvise on a public site. Use the printed log as the demo.
  • Pre-cut tool cards. Include a bright Halt card in every deck.
  • Write two task prompts on the board: Defined (campsite constraints) and Fuzzy (plan a cool weekend).

Instructional sequence

Answer vs. do

5 min
  1. Read two outputs. A: a paragraph about how to reserve an Idaho state park site. B: a log that searched dates, compared three sites, and paused for a human to confirm.
  2. Students label A Chat and B Agent on the packet.
  3. Ask what B did that A did not, and what B still must not do.

Goal, plan, tools, check, loop

10 min
  1. Write the five-part skeleton. Defined task goes above it. Guardrails go beside it. Human halt sits at the end of every loop.
  2. Walk the printed campsite log. Circle each tool call. Cross out the illegal Submit the agent almost made.
  3. Contrast a fuzzy task ('plan a great weekend') that an agent cannot finish because success is undefined.
  4. Name Idaho twins: county password reset, hospital checklist, irrigation draft, HVAC parts search. All need a named approver.

Label the log

8 min
  1. As a class, color-code the second log: goal, plan, tool, result, human gate.
  2. Find the step where the agent exceeded the guardrail (spent money, emailed a patient, filed a permit).
  3. Rewrite the guardrail in eight words or fewer.

Defined or not?

8 min
  1. Students sort six task slips: county records request, 'help Grandma,' irrigation set points for field 4, 'fix the mill,' hospital discharge checklist, 'make practice fun.'
  2. For each defined task they add a guardrail. For each fuzzy task they write what would have to be specified.
  3. They star one task that should stay fully human even if it could be specified (for example, telling a family a diagnosis).

Real-world examples

  • Idaho Parks: an agent can search open campsites and list fees; a human confirms fire restrictions and hits reserve.
  • County IT: an agent can triage ticket categories; a technician still resets access for a deputy or a clerk.
  • Hospital: an agent can assemble a discharge checklist from a template; a nurse still teaches the patient.
  • Agriculture: an agent can draft pivot run times from weather data; the grower still walks the field for a plugged nozzle.
  • Trades: an HVAC agent can find a compatible part number; the technician still verifies the unit tag before ordering.

Hands-on activity

Tabletop agent, Halt card required

14 min
  1. Groups of four: Agent, Tool, Human approver, Auditor. They draw a defined Idaho task from the packet.
  2. The Agent may only act by playing a tool card. The Tool returns a printed result slip (including one error slip in the deck).
  3. After two loops the Agent must request Halt. The Human either approves a draft, demands another check, or shuts the task down.
  4. The Auditor writes the log: goal, steps, tools, error, final human decision. No group is allowed to 'Submit' without the Halt card.
  5. Teacher projects a live agent demo only on a district-approved tool with a harmless search. If none exists, the printed log is the demo. Chromebooks are optional note sheets, not required.

Discussion questions

  1. What made a task defined enough to attempt?
  2. Which tool card was most dangerous if the human fell asleep?
  3. How is an agent different from a dishwasher, a thermostat, or a chatbot?
  4. Should a school agent ever email families without a staff click? Why or why not?
  5. What error in the deck was most like a hallucination from ECT.5, and who had to catch it?

Differentiation

Support

  • Pre-label the first two log lines. Give a script for the Agent role.
  • Limit the simulation to one loop plus Halt.
  • Offer a word bank: goal, tool, result, approve, stop.

Challenge

  • Add a second agent that must hand off to the first, and write the interface a human still owns.
  • Design a guardrail for money, medical, or legal actions in one sentence each.
  • Explain why AP.3 (constructing agents) will need this recognition first.

Multilingual learners

  • Act out tool, halt, and approve with gestures before the simulation.
  • Allow the auditor log in bilingual bullets.
  • Pre-teach agent vs. chatbot with the warmup pair of outputs.

IEP / 504

  • Assign Auditor or Human if rapid talk is a barrier; still require one spoken Halt.
  • Provide a large Halt card and a one-page log template.
  • Allow a quieter pair simulation instead of four roles.

Assessment

Formative

  • Warm-up Chat vs. Agent labels.
  • Color-coded logs and the defined/fuzzy sort.
  • Auditor sheets from the tabletop run.

Summative

  • Exit ticket: In 6–8 sentences, explain how an AI agent accomplishes a defined Idaho task, using goal, tool, check, and human halt. Include one task you would not delegate.
  • Packet log scored for correct labels and a refused Submit.

Success criteria

  • I described an agent as a loop with tools, not as a person.
  • I used a defined task with a finish line and a guardrail.
  • I required a human halt before send, spend, or file.
  • I named a task that should stay human even if a system could attempt it.

Responsible use, ethics, and privacy

Responsible use

Simulate on paper. Do not connect student accounts to third-party agents that can send mail or spend. If a district tool can take actions, the teacher holds the login and the class watches a harmless search only.

Ethics

Action raises the stakes above chat. An agent that files, alerts, or orders can harm people faster than a wrong paragraph. Defined tasks, guardrails, and a named human are ethical requirements, not extras. Do not treat the agent as a coworker who 'meant well.'

Privacy

FERPA: no student PII in unapproved tools. Agents that can see inboxes, calendars, or student information systems are high risk. Do not paste names, IDs, schedules, or health data into unapproved agent tools. Use fictional tickets in the packet. School accounts only, and only if the district has approved that agent.

Reflection

  1. Where did our table want the agent to 'just finish,' and why was that a trap?
  2. What guardrail would I demand at a job I actually want?
  3. How will I tell a chatbot from an agent the next time a product page uses both words?

Homework

On paper, write one defined task an agent could attempt in an Idaho job you know (farm, shop, clinic, park, county office, home). List the goal, two tools, one check, one guardrail, and the human who must Halt. Then write one nearby task that must stay human. No live agent accounts.

Closing

Hold the Halt card up as a class. Repeat: defined task, tools, check, human. Collect auditor logs. Point forward to AP.3, where they may construct a tiny agent only after they can recognize what one is actually doing.

Extensions and cross-curricular links

Go further

  • 90-minute block: run a second simulation with a planted tool error and write a post-incident note from the Auditor.
  • CTE: map an HVAC or ag workflow into steps an agent might draft vs. steps a licensed person must perform.
  • CS preview: storyboard an agent without writing code; AP.3 will build later.
  • Civics: draft a three-rule policy for a county chatbot that is not allowed to become an agent on permits.
Computer Science
Algorithms as steps with a stop condition; agents add tools and loops around those steps.
Business / Personal Finance
An agent that can spend is a debit card with a language model. Guardrails first.
Health Sciences
Checklists can be assembled by a system; teaching, consent, and discharge remain human.
Government
Public agencies may use agents to sort, not to issue licenses or deny benefits without a person.