Five-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Step 5 needs a person. Step 5 is waiting for approval right now. Click it to open the full walkthrough.
Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.
How an Early-Stage EdTech Company Built Live Tutoring That Hands Instructors the Children Who Need Them
Every child's practice, from the minute the day's work is picked to the minute an instructor decides who needs them. The agents plan each child's session off their own record, tutor them through it one step at a time, grade every answer on the spot, and redraw the map of what the child knows. An instructor is handed six children with the evidence, not two hundred and forty with a gradebook. Five stations, run as a loop: what the instructor decides at station 5 changes what station 1 picks for that child tomorrow.
What Was Broken
What We Built
Five stations and 11 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Lesson Selection
Nobody is asked anything hereAgents pull each child's record, find the gaps that block the next skill, and write a plan that reviews what is fading before it adds anything new. Nothing waits on a person here. A child whose placement never finished gets no guessed plan, they get held.
Live Tutoring
Nobody is asked anything hereAgents explain each stuck step in the child's own words, move the difficulty with the child's answers, and end a session early rather than push a child who has stopped trying. Nothing waits on a person here. A question off the lesson is never answered, it is queued for the instructor with its transcript.
Answer Grading
Nobody is asked anything hereAgents score every answer while the child is still looking at it, name the exact step where a miss happened, and send an answer no rubric covers to the instructor rather than guessing a grade. Nothing waits on a person here. Feedback the agents cannot tie to the child's own work is withheld, and the worked example shown instead.
Progress Mapping
Nobody is asked anything hereAgents redraw each child's skill map from the graded work, close a gap only when it holds across sessions, and queue any whole-level move for the instructor rather than making it alone. Nothing waits on a person here. A level move is never made by an agent, it waits at the instructor's desk with the sessions behind it.
Instructor Desk
A person answers hereAgents watch every child's engagement and completion across sessions, flag the ones slipping with the evidence attached, and draft an intervention note for each. An instructor decides every intervention, clears every level move, and reads every queued exchange. This is where a person works. Everything else exists so that six children, not two hundred and forty, reach this screen.
Where a Person Decides
Operating Model
This changes how work flows through the team.
What Transfers, What Must Be True
Failure Modes
What breaks this pattern:
The child gets new material that sits on something they never learned. Every lesson after that teaches them to fail, and the plan calls it progress.
A child asks something outside the lesson and the tutor improvises. No instructor ever sees what the child was told, so nobody can stand behind it.
The agent scores an answer no rubric covers. The child learns the wrong lesson twice: once from the mistake, once from the grade that missed it.
The agent drops a child a level on its own read of one bad week. The first person to notice is the parent, and nobody can say who decided.
Directional Outcomes
What the agent counts, and the station that counts it.
These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.
What Runs Where
Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.
pick-next-workstudents/<id>/plan/<date>.jsonplacement-unfinishedstudents/<id>/plan/**, the record and the map read-onlyrun-sessionstudents/<id>/sessions/<date>.jsonoff-lesson-exchangestudents/<id>/sessions/**, the plan read-onlygrade-and-feedbackstudents/<id>/grades/<session-id>.jsonno-rubric-for-answerstudents/<id>/grades/**, the session record read-onlymap-progressstudents/<id>/map.json, students/<id>/path.jsonlevel-movestudents/<id>/map.json, students/<id>/path.json, the grades read-onlybuild-the-deskdesk/<date>/<child-id>.jsonawait-instructor-decisiondesk/**, everything else read-onlyStack
Every system this agent reads or writes.