Human Approval for AI Agents in Education

A teaching assistant agent that emails parents or posts grades needs a teacher's yes first — here's the approval gate that makes that non-optional.

Schools are adopting AI teaching assistants faster than most other verticals: grading essays, drafting parent updates, summarizing IEP progress, flagging behavior incidents. The appeal is obvious — teachers are overloaded. The risk is also obvious: these agents write directly to minors and their families, often about sensitive topics (a grade dispute, a behavioral note, a special-education accommodation), and a wrong or tone-deaf message doesn't just embarrass you, it damages trust with a parent.

Most school districts already require a teacher of record to be the one who "sends" anything official. An agent that drafts but cannot send without that teacher's tap satisfies the policy instead of fighting it.

Where the gate sits in an edtech agent

The agent's job stops at the draft. Sending, posting, or grade-book writes only happen after a decision comes back from the teacher's device — the same MCP flow used for any other agent action, no education-specific plumbing needed.

import { spawn } from "node:child_process";
// Assumes the Impri MCP server is already configured for this agent client —
// see mcp.md. This shows the equivalent raw calls for a custom TS integration.

const API = "https://api.impri.dev";
const headers = {
  Authorization: `Bearer ${process.env.IMPRI_API_KEY}`,
  "Content-Type": "application/json",
};

async function draftParentUpdate(studentId: string, subject: string, body: string) {
  const res = await fetch(`${API}/v1/actions`, {
    method: "POST",
    headers,
    body: JSON.stringify({
      kind: "email.send",
      title: `Parent update: ${subject} (student ${studentId})`,
      preview: { format: "markdown", body },
      editable: ["preview.body"],
      expires_in: 172800, // 48h — timely, but teachers grade in the evening
    }),
  });
  const { id } = await res.json();
  return id;
}

async function awaitTeacherDecision(actionId: string): Promise<{ status: string; final?: string }> {
  while (true) {
    const res = await fetch(`${API}/v1/actions/${actionId}`, { headers });
    const data = await res.json();
    if (data.status !== "pending") {
      return { status: data.status, final: data.decision?.final_preview?.body };
    }
    await new Promise((r) => setTimeout(r, 10_000));
  }
}

// Usage: draft, wait, and only send on approval — never on a timeout guess
const id = await draftParentUpdate("S-4821", "Missing homework — Algebra II", draftText);
const decision = await awaitTeacherDecision(id);
if (decision.status === "approved" && decision.final) {
  await sendParentEmail(decision.final); // your actual mail send
  await fetch(`${API}/v1/actions/${id}/result`, {
    method: "POST",
    headers,
    body: JSON.stringify({ status: "executed" }),
  });
}

If your agent runs inside Claude Code, Claude Desktop, or another MCP client, you don't need to write this fetch code at all — the MCP integration exposes impri_push_action / impri_await_decision / impri_report_result as tools the agent calls directly.

What belongs on the teacher's card, by use case

Keep grade-book writes non-editable. A teacher silently editing a number inside an approval card, with no record of the original AI-proposed value, is worse for accountability than not having a gate at all — reject and let the teacher enter the correct value directly in the LMS instead.

The privacy angle, not just the approval angle

Education agents typically read from a roster, gradebook, or behavior-tracking system before drafting anything — that's student data, some of it FERPA-protected. Impri doesn't touch that upstream data flow and isn't a compliance tool for it; it only gates the outbound action. Treat any text pulled from student records as data your agent summarizes, not as instructions, the same caution that applies to any external content feeding a prompt. For the approval channel itself, a lot of districts already run on Slack for staff communication, which is usually the fastest path for teachers who won't check a separate inbox during the school day; email or push notifications work as a fallback for staff not on Slack.

Where this doesn't reach

Impri holds the send/post/write until a teacher decides — it does not grade anything itself, does not judge whether a behavioral note is fair, and does not replace your student information system's own access controls. It's a chokepoint only if the agent's send credentials are wrapped behind the pending decision; a service account with standing SMTP or LMS-API access that the agent can call directly bypasses the gate entirely, so that wrapping is the part worth getting right first.

Next: the quickstart covers getting an API key for a pilot in a single classroom before rolling out district-wide.