A connected agent gets a fixed set of tools. It can read your family's learning state and propose work, but a parent stays in the loop for anything the student would see. This is the full picture.

Reads

These tools return information and change nothing:

  • Family context: students, subjects, school years, upcoming evaluations, recent results and class summaries.
  • Subjects with their curriculum topics.
  • Pending drafts and pending assignment requests, so it knows what you asked for and what is still waiting on you.
  • Submitted assignments and a submission with its photos, so it can suggest a correction.
  • Reviewed correction drafts, correction edits and recent rejections, so it can learn from what you accepted, changed or refused.
  • The student learning log and top weak topics.

Writes that wait for you

  • Draft an assignment: title, instructions, subject, optional due date and a rationale for you. Lands in the approval queue.
  • Revise a draft after you reject it or leave a note. There is a cap on revisions per draft.
  • Suggest a correction for a submission: score, feedback, weak topics, rationale, and optionally a paired follow-up assignment. Lands on the correction page for you to use or dismiss.
  • Approve a draft: only if you ask the agent to, and only as you. The audit log records that it happened through the agent.

Writes that land directly

These are records, not student-visible work, so they do not go through the queue:

  • Create a subject and replace a subject's topics.
  • Record a class summary, an evaluation or a test result.

All of them appear in Family → Activity.

What an agent cannot do

  • Make anything visible to a student without a parent approval, unless the family turned on auto-approval.
  • See or touch another family. Tools resolve students and subjects inside your family only.
  • Pair devices, invite members, change settings or delete anything.
  • Reach the Student app. There is no channel from an agent to the device.

The quality gate

Every draft and suggestion is checked before it is stored as pending: minimum length for title, instructions and rationale, no AI meta-commentary in the student-facing voice, no student full names, no duplicate weak topics. Failures are stored as rejected so you can see what was tried, and the agent gets the list of violations back so it can fix and retry.

Rate limits

Draft, revise, suggest and topic-replacement calls are limited per family and per tool, by default 10 per minute. An agent that goes over gets a retry-after and the event shows as Agent throttled on the Dashboard and under Family → Trust. Reads are not limited.

Seeing what it did

  • Family → Activity lists every tool call with the tool name and outcome, filterable by week and by tool, and exportable as CSV.
  • Family → Trust shows reliability over the last four weeks: the share of reviewed drafts you approved, per-tool counts and failures this week, and recent throttle events.
  • The weekly digest counts drafts approved and rejected and corrections shipped.