Skip to content

AI Agent Loop Explorer

Run the AI Agent Loop Explorer MicroSim fullscreen

About This MicroSim

Four stages: input, draft, human review, send or reject. The sim steps through them with the actual text at each one, because the argument only works with real content in front of you.

Every draft here contains exactly one wrong detail, and it always reads perfectly. A start time carried over from last term's calendar entry. A room number with its digits transposed. A $20 fee that does not exist, invented because most clubs have one. None of these drafts is badly written — they are fluent, correctly formatted, in the club's voice, and wrong in one place.

That is the failure mode worth teaching. A badly written draft gets caught by anybody. A fluent one gets skimmed and sent, and the wrongness only surfaces when thirty families arrive at a locked library half an hour early.

Press "What if no one reviewed this?" on each scenario. The parent-chatbot one is the worst of the three: a family that could not pay the invented $20 quietly does not reply, and nobody ever finds out why. That is a failure with no error message attached.

Stage 3 is the one people want to remove, because it is the only stage that is not automatable and therefore the only one that still costs somebody's time. It is also the only stage that catches a fluent mistake. Those two facts are the same fact.

Reject works the same way as send — it is the same stage saying no.

How to Use

  • Press Next to move through the four stages. Read the draft at stage 2 carefully before moving on; the wrong detail is findable.
  • At stage 3 the correction is shown on its own, so it is clear how small it is.
  • Press What if no one reviewed this? at any stage for the unreviewed version and what it costs.
  • Switch Scenario for two more, including the one where the agent invents a fee out of nothing.

Lesson Plan

Bloom level: Understand (L2) -- explain

Learning objective: Given a scenario, the learner explains what happens at each stage of an AI agent's loop and why the human-review stage cannot be skipped.

Before the sim (5 min). Read the registration reminder draft aloud without saying anything is wrong with it, and ask whether it is ready to send. Most rooms say yes.

With the sim (15 min). Walk all three scenarios. The parent chatbot is the one to end on, because the invented fee is not a typo — the agent produced a fact that was never in its input, and that is a different and harder problem than a wrong number.

After the sim (20 min). Write the club's actual review rule: who reviews what, before what goes out. "Anything with a time, a place, or a price in it" is a good first draft of that rule.

Check for understanding. Ask: "The agent has been right twenty times in a row. Can you stop reviewing?" No — the drafts were always fluent, so twenty correct ones tell you nothing about the twenty-first. Being right often is what makes an unreviewed agent dangerous rather than safe.

Embedding This MicroSim

Paste this into any page of the book, adjusting the relative path to docs/sims/ for the page's depth:

<iframe src="../../sims/ai-agent-loop-explorer/main.html" width="100%" height="562" scrolling="no"></iframe>

Specification

The full specification below is extracted from Chapter 32: AI Agents for Registration, Scheduling, and Communication.

Type: microsim
**sim-id:** ai-agent-loop-explorer<br/>
**Library:** p5.js<br/>
**Status:** Specified
**Template:** https://github.com/dmccreary/ai-strategy-for-education/tree/main/docs/sims/ai-agent-anatomy

Purpose: Let a learner step through the four-stage loop every AI agent in this chapter follows -- input, draft, human review, and send or reject -- using the registration-reminder scenario from the worked example above as the default walkthrough.

Bloom Taxonomy: Understand (L2)
Bloom Taxonomy Verb: explain

Learning objective: Given a scenario, the learner explains what happens at each stage of an AI agent's loop and why the human-review stage cannot be skipped.

Instructional Rationale: This is an Understand-level objective, so the sim uses a step-through with concrete data at each stage rather than a continuous animation -- a learner needs to see the actual draft text and the actual correction made, not watch a generic process play out with no real content to reason about.

Canvas layout:
- Left (450px): the current stage's content -- the input goal, the agent's draft, the reviewer's markup, or the final outcome
- Right (150px): a four-step progress tracker (Input -> Draft -> Human Review -> Send/Reject) with the current step highlighted

Data Visibility Requirements:
  Stage 1 (Input): Show the plain-language goal typed in -- "Draft Tuesday's session reminder"
  Stage 2 (Draft): Show the agent's full draft text, including the incorrect start time it carried over from an old calendar entry
  Stage 3 (Human Review): Show the reviewer's markup -- the corrected start time circled, with a short note "fixed time, kept the welcome-back line"
  Stage 4 (Send/Reject): Show the final, corrected message as it actually went out, side by side with the original draft so the one changed detail is visible

Interactive controls:
- Dropdown: choose a scenario (Registration Reminder, Room-Change Announcement, Parent Chatbot Reply) -- each loads its own four-stage content
- Button pair: "Next" / "Previous" to move through the four stages
- Button: "What if no one reviewed this?" -- reveals a short callout showing the uncorrected draft (wrong start time) as it would have looked if sent without review

Default parameters: Scenario = "Registration Reminder"; starts at Stage 1

Implementation notes: p5.js canvas with createSelect for the scenario dropdown and createButton for Next/Previous and the "what if" button; each of the three scenarios' four-stage content lives in a small lookup object so switching scenarios mid-walkthrough resets cleanly to Stage 1.

References