Skip to content
Idaho AI Classroom
All lessons
AI FluencyApplication9-12.AIF.AP.3

Build a Narrow Agent

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

Standard quoted exactly

Construct AI agents to accomplish a defined task.

Student-friendly learning targets

  • I can construct a narrow AI agent by specifying a goal, allowed tools, constraints, and success checks on a worksheet, without having to write Python.
  • I can choose a defined task small enough that a human can verify every step, such as finding three reputable sources on Idaho water rights or drafting a packing list for a 10th-grade Sawtooth field study given a forecast.
  • I can run the spec (in a district-approved agent builder if one exists, or as a human role-play of the agent) and mark each success check pass or fail.
  • I can explain why a narrow agent with a human in the loop is safer than a vague instruction like 'do my research' or 'plan the whole trip.'

Essential questions

  1. What does 'construct' mean if I am not writing a program?
  2. How narrow does a task have to be before an agent can finish it and a human can check it?
  3. Which tools should this agent be allowed to use, and which must it be forbidden to use?
  4. Who is responsible if the agent lists a URL that does not exist or packs a tent and forgets water?

Objectives

  1. Students will complete an Agent Spec worksheet: task name, one-sentence observable goal, inputs, allowed tools, forbidden tools, constraints (including no PII), success checks, human verification steps, failure modes, and stop conditions.
  2. Students will construct a spec for one of two defined tasks: (A) find three reputable sources on Idaho water rights and list them with URLs to verify, or (B) draft a packing list for a 10th-grade field study in the Sawtooths given a weather forecast.
  3. Students will execute the spec through a district-approved agent builder if available, or through a partner role-play that may use only the tools the spec allows.
  4. Students will verify every step as a human: open each URL, or check each packing item against the printed forecast and school field-trip rules.
  5. Students will record at least one failure or hallucination (broken URL, missing layer, invented agency) and one change they would make to the spec.

Key vocabulary

Agentic AI
AI systems designed to independently plan, execute, and adjust actions to achieve a specific goal with minimal human intervention, often capable of working across multiple tools or applications.
Agent (in this lesson)
A narrow worker you design for one defined task. You construct it by writing its goal, tools, constraints, and success checks. You do not have to code it in Python.
Constraint
A rule the agent must not break, such as 'use only .gov or .edu sources,' 'no student names,' or 'pack for 24 hours, not a week.' Constraints keep a narrow agent from wandering.
Success check
An observable test a human can apply when the agent finishes: 'three URLs that open and name Idaho water rights,' or 'every clothing item matches the forecast overnight low.'
Human-in-the-loop
A model of interaction where a human is required to participate in a process, such as verifying, approving, or correcting, to ensure the accuracy, safety, and ethical alignment of AI-generated outcomes.
Human-Centered
An approach where AI is used specifically to support human inquiry. It ensures that humans remain responsible for the initial objective, final decision-making, oversight, and ethical reflection at the beginning and end of every task (Human AI Human).

Teacher background

Idaho standard 9-12.AIF.AP.3 is one sentence: Construct AI agents to accomplish a defined task. For grades 9–10, 'construct' does not mean ship a Python agent with API keys. It means specify the agent so tightly that another person (or a district-approved builder) could follow it: goal, tools, constraints, success checks, and a human who verifies every step. That reading matches the glossary terms Agentic AI, human-in-the-loop, and Human-Centered (Human AI Human). Keep tasks narrow. Two classroom-ready jobs work in Idaho without collecting PII: (1) find three reputable sources on Idaho water rights and list them with URLs a human will open; (2) draft a packing list for a 10th-grade field study in the Sawtooths given a weather forecast the teacher prints. Reputable water-rights starting points a teacher can verify include the Idaho Department of Water Resources, Idaho statutes or the Attorney General's water-law materials, and University of Idaho Extension or Idaho Law Review explainers — not a random blog and not a URL the model invented. Sawtooth packing must obey school trip rules (no personal vehicles, medications handled by staff, no real home addresses in the prompt). If the district has not approved an agent builder, the worksheet plus a partner who role-plays the agent is the construction. Do not send students to unpaid public agent sites that scrape chats. This lesson sits with AP.2 in Weeks 10–14; keep the agent as small as yesterday's script, and keep a human on every step.

Materials and prep

Materials

  • Agent Spec worksheet with labeled boxes: task name, goal, inputs, allowed tools, forbidden tools, constraints, success checks, human verification steps, failure modes, stop conditions.
  • Two task cards. Card A: 'Find three reputable sources on Idaho water rights and list them with URLs to verify.' Card B: 'Draft a packing list for a 10th-grade field study in the Sawtooths given a weather forecast.'
  • Printed forecast for Stanley or the Sawtooth National Recreation Area (public NWS text, no student locations). Printed field-trip rules excerpt (school-provided, no student medical information).
  • Printed 'starter map' of water-rights sources for the teacher (IDWR, a .gov statute page, UI Extension). Do not hand this to students until after the first verification so they still have to check.
  • Optional: district-approved agent builder or a locked teacher account. If none, skip without penalty.
  • Role-play badges or sticky notes labeled AGENT and HUMAN VERIFIER.
  • Verification log: step, agent output, human check (opened URL / compared to forecast), pass/fail, notes.
  • PII and FERPA card: the agent may not collect student names, home addresses, phone numbers, or health information. Packing lists use 'a 10th grader' not 'Maria from McCall.'
  • Offline packet: a filled sample spec with two planted errors (a 404-style fake URL; a packing list that omits water) for students to catch.

Before class

  • Print Agent Spec worksheets and both task cards. Decide whether an approved builder exists. If not, plan role-play as the default, not as a downgrade.
  • Pull a current public forecast for Stanley, ID or Sawtooth NRA and a one-page field-trip rule sheet. Redact any real student information if you are adapting a past trip form.
  • Walk both tasks yourself. For water rights, confirm three URLs that actually open. For packing, pack against the forecast overnight low and a 24-hour trip. Plant one error in the sample packet.
  • Post the construction definition: goal + tools + constraints + success checks + human verification. Post what this lesson is not: 'write Python,' 'let the agent do my homework,' 'give it my address.'
  • Arrange pairs: one constructs, then they swap verifier roles. Solo students verify against the printed forecast or the teacher starter map.
  • 90-minute block: keep this sequence. Add about 20 minutes so every spec is executed and every URL or packing item is checked, plus about 15 minutes of discussion on failure modes (hallucinated agencies, over-packing, missing water). Do not widen the task to 'plan the whole trip' or 'write the research paper.'
  • Coordinate with the librarian if possible: reputable source means the human can name the publisher, not that the chatbot said 'reputable.'

Instructional sequence

Recipe versus wanderer

5 min
  1. Show two instructions. Wanderer: 'Do my research on water and also plan our mountain trip.' Recipe: 'Return three sources on Idaho water rights. Each must have a URL. I will open each URL. Stop after three.'
  2. Students label which one is a defined task an agent might finish and a human might check in one class period.
  3. Name the day's claim: we will construct a narrow agent on paper. Construct means specify. It does not require Python. A human verifies every step.
  4. Point at the two task cards. Both are Idaho work. Both are small on purpose.

What 'construct' looks like on a worksheet

10 min
  1. Read the standard aloud exactly: 'Construct AI agents to accomplish a defined task.' Write construct on the board and unpack it: goal, tools, constraints, success checks, human verification, stop conditions.
  2. Read the glossary sense of Agentic AI and then narrow it for 9–10: we are not launching an unsupervised worker across the open internet. We are writing a job description a tool (or a partner) must obey.
  3. Model Task A on the document camera. Goal: 'List three sources on Idaho water rights, each with a URL I can open today.' Allowed tools: search of .gov and .edu only, or the school library page. Forbidden: social media as the only source, invented citations, student names. Success checks: three URLs; each opens; each page is clearly about Idaho water rights. Human verification: I click each link. Stop: after three verified sources or after five minutes, whichever comes first.
  4. Show a bad spec: 'Get me everything on water and make it sound smart.' Ask why a human cannot verify it in this period.
  5. Show the Human AI Human loop on the board: human writes the goal, agent (or role-play) works, human checks, human decides whether the job is done.
  6. FERPA reminder: packing-list prompts use a generic 10th grader and a public forecast. No home addresses, no medication lists, no classmate names.

Co-write one spec, then try to break it

10 min
  1. As a class, fill the Agent Spec boxes for Task A or B (teacher choice; use the other as independent practice). Fill every box; empty boxes are not allowed.
  2. Invite a student to play AGENT. The agent may use only allowed tools: for water rights, the teacher may permit a search on a projector or a printed IDWR snippet; for packing, the agent may read only the printed forecast and the rule sheet.
  3. A second student plays HUMAN VERIFIER and uses the verification log. If a URL is not opened, it does not count. If a packing item ignores the overnight low, it fails.
  4. Class names one constraint to add after seeing a miss (for example: 'Do not list Facebook groups as sources' or 'Include water and a warm layer no matter what').
  5. Students now choose the other task, or the same task with their own wording, for independent construction. They may not skip the worksheet and 'just chat.'

Write the spec; do not skip to the toys

12 min
  1. Each student (or pair) completes a full Agent Spec for Task A or Task B. Complete sentences in the goal box. Tools listed by name. At least three constraints, one of which is 'no PII.' At least three success checks a human can observe.
  2. Peer desk-check: partner tries to find a way the spec would let the agent wander (for example, packing for a week, citing Wikipedia only, asking for the student's street address). Author revises before any execution.
  3. If a district-approved builder is available, students may paste the spec into it after the desk-check. If not, they stop at a complete worksheet and wait for the hands-on role-play. Both paths meet the standard.
  4. Students with no partner use the planted-error sample packet as the agent output and complete the verification log against it.
  5. Teacher circulates for missing stop conditions. A spec without a stop will run until the period ends; make them write one.

Real-world examples

  • A county emergency manager in eastern Idaho might use an agent to draft a reverse-911 outline from a public weather feed; a human still records the message and sends it.
  • A school librarian constructs a search agent that may return only .gov and .edu sources on water rights; the librarian still opens the links before they go on a research guide.
  • An outdoor-education teacher at a McCall or Stanley field study uses a packing-list agent against an NWS forecast, then checks the list against district trip rules and a real duffel.
  • A 4-H robotics team writes a narrow agent spec for 'sort by size only'; a human still pulls damaged produce the robot cannot see.
  • Idaho Department of Water Resources staff, not a chatbot, remain the authority on water rights; an agent that invents a decree number has failed even if the sentence sounds official.

Hands-on activity

Execute, then human-verify every step

8 min
  1. Pairs assign roles: AGENT follows the spec exactly (builder or role-play). HUMAN VERIFIER opens URLs or checks packing items against the printed forecast and trip rules. Switch after four minutes if time.
  2. Fill the verification log. Unopened URLs are automatic fails. Packing lists that omit water, a warm layer when the overnight is below 40 F, or that add a personal vehicle against trip rules fail.
  3. Mark hallucinations in the margin: a source that does not exist, a misnamed agency, a forecast the agent invented instead of reading. One hallucination is enough to require a spec revision.
  4. Each student writes one spec change: add a constraint, tighten a success check, or narrow the goal. This is the construction improvement, not extra code.
  5. Block extension (90 minutes): run the revised spec once more. Then a short seminar: could this agent be allowed to send an email, book a van, or file a water-rights form? (The answer you want is no — those are human actions.) Optional approved-builder demo by the teacher only, on a projector, with the class calling stop conditions out loud.

Discussion questions

  1. Where in your spec was the agent still free to wander, and what constraint closed that hole?
  2. Why does 'three URLs I will open' beat 'good sources' as a success check?
  3. If the agent packs a stove and the trip rules forbid student stoves, who is at fault — the tool or the author of the spec?
  4. What would go wrong if we let an agent 'just handle' water-rights research for a graded paper?
  5. How is this construction different from yesterday's tiny generated script, and how is the human's job the same?
  6. When should we not build an agent at all and just do the task ourselves?

Differentiation

Support

  • Provide a partially filled spec with the goal and one constraint already written; the student adds tools, success checks, and verification steps.
  • Offer a word bank of tools (school library page, printed forecast, IDWR site, calculator) and of forbidden tools (social media only, invented URL, home address).
  • Let the student construct Task B (packing list) if reading dense .gov pages is the barrier; verification is then tactile (forecast numbers, clothing items).
  • Scribe the worksheet from the student's oral spec; the thinking is the standard, not the handwriting.

Challenge

  • Add a second success check that catches a known hallucination type (URL that does not mention Idaho, packing list that ignores wind).
  • Write a stop condition that would prevent the agent from starting a new task it was not given.
  • Compare two specs for the same task and argue which one a substitute teacher could run tomorrow without you in the room.

Multilingual learners

  • Worksheet boxes may be completed in the student's strongest language; the three URLs and the packing items stay in a form the verifier can check.
  • Pre-teach goal, tool, constraint, verify, and stop with a bilingual card.
  • For Task A, allow one source in Spanish if it is still an official or extension publisher the student can identify, plus two English .gov/.edu sources, or the reverse, as long as the human can verify each.

IEP / 504

  • Fewer boxes required for credit: goal, two tools, two constraints, two success checks, and a named human verifier. Same verification rigor.
  • Extended time to open URLs in the library lab. Printed pages instead of live links if screens are a barrier.
  • Role-play only; no requirement to use an agent builder.
  • Quiet instructions and a checklist with boxes to tick rather than long prose. Adult may read the forecast aloud.

Assessment

Formative

  • Warm-up label of recipe versus wanderer.
  • Desk-check of the spec before execution; teacher initials that no PII box is empty.
  • Verification log in progress: at least one URL opened or one packing item checked in class.

Summative

  • Completed Agent Spec with every box filled, plus a verification log that shows a human check on each success criterion.
  • One written spec revision after a miss or a planted error, and an exit ticket: 'Name the human's job at the start and at the end of this agent.'

Success criteria

  • The task is defined and narrow; it matches Card A or Card B or an equally small teacher-approved variant.
  • The spec includes goal, tools, constraints, success checks, and a stop condition. Construction is visible on paper even if no builder was used.
  • A human verified every step. Unchecked output does not count as success.
  • No PII was entered. No classmate's information was used. The student can name one failure mode.

Responsible use, ethics, and privacy

Responsible use

Construct on the worksheet first. Use a district-approved agent builder only if one exists; otherwise role-play. Do not create accounts on public unsupervised agent sites. Do not give an agent permission to send email, spend money, post to the web, or contact agencies. The human starts the task, checks every step, and stops the agent. Credit the builder if used. Offline role-play is full credit. Specs stay in the classroom system or on paper.

Ethics

A narrow agent can save time on fetching links or listing gear. It cannot take responsibility for a broken source, a student who goes to the Sawtooths underdressed, or a water-rights claim that is wrong. Writing a vague spec and blaming the tool is an ethical failure of the constructor. Inventing URLs or packing without reading the forecast is a hallucination the human must catch. Do not build agents for tasks that should stay fully human: grading peers, medical decisions, reporting a classmate, or contacting government about a real water right in a student's family name.

Privacy

FERPA and good sense: no student names, home addresses, phone numbers, health or medication details, or family water-right owner names in the spec or the prompt. Use 'a 10th grader' and a public forecast. Do not upload a real permission slip. Agent builders, if used, must be district-approved with a data-privacy agreement. Role-play on paper never leaves the room. Verification logs that include a student name are education records and stay in the district grade system, not in a public chatbot.

Reflection

  1. Which box on the spec did the most work — goal, constraint, success check, or stop — and what would have happened if I had left it blank?
  2. What did the human verifier catch that the agent (or role-play partner) missed?
  3. Would I let this agent talk to the web unsupervised overnight? Why or why not?
  4. How will I use Human AI Human on the next research or trip-planning job in another class?

Homework

Revise one constraint and one success check on your spec using tonight's thinking. If you chose water rights, try to open your three URLs at home or at the public library and write pass/fail; if a link fails, find a replacement by hand, not by trusting a new unverified list. If you chose packing, check the next day's Stanley or Sawtooth forecast (radio, printed paper from the teacher, or a family device) and note one item you would add or drop. Students without internet receive a printed forecast and a printed IDWR page as the homework packet. Do not sign up for new agent websites. Do not include your address or a family member's name.

Closing

Repeat the standard: construct AI agents to accomplish a defined task. Point at a completed worksheet: that is construction. Point at a verification log: that is the human. Collect specs and logs. Two students name one hallucination they caught. End with the loop on the board: human writes the goal, the narrow agent works, the human checks, the human decides. Preview how this skill returns in later standards when agents appear inside real applications: the job description you wrote today is still the safety rail.

Extensions and cross-curricular links

Go further

  • Librarian station: convert a successful water-rights spec into a one-page research guide other classes can use, still with the instruction to open every URL.
  • Outdoor club: compare the agent packing list to what a Sawtooth ranger or a school trip lead actually packed last year (guest or recorded interview).
  • CS extension (not required): if a district builder exists, save the spec as a reusable template with the teacher's name as owner, not the student's personal account.
  • Write a third task card for a local job (greenhouse watering check, FFA meeting agenda reminder) that is still verifiable in ten minutes.
Civics / Idaho history
Water rights are a live Idaho policy topic. Students practice finding primary publishers (IDWR, statute) instead of accepting a chatbot's summary.
Earth science / outdoor education
Field-study packing against a real forecast in the Sawtooths; weather numbers beat generic 'bring a jacket' advice.
English / research
Source evaluation: a URL is not a source until a human opens it and names the publisher.
Health / PE
Trip safety constraints (water, layers, school rules) belong in the spec so an agent cannot optimize them away.