Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 5 and 6 need 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 a Growth-Stage Technology Company Built Learning Paths Where Difficulty Moves with the Learner
Every learner, from the day they enrol to the day they finish, on a path built from what they actually know. The agents place each learner, build their path, move the difficulty with them, answer their stuck questions, grade their work the same day, and put on an instructor's desk only the learners whose next step carries a person's name: an outreach message, a borderline final, a change to the course itself. Six stations, run as a loop: the step that stalled eleven learners this week is the module station 2 builds paths around next week.
What Was Broken
What We Built
Six stations and 12 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Learner Placement
Nobody is asked anything hereAgents give every new learner an adaptive placement, read the answers into held, shaky and missing skills, and hold any placement whose answers contradict each other. Nothing waits on a person here. A held placement is never averaged into a starting level; it reaches the instructor's desk with its answers attached.
Path Building
Nobody is asked anything hereAgents build each learner's path from the modules that teach their missing skills, size it to the hours they gave, and flag any gap no module covers instead of assigning the nearest one. Nothing waits on a person here. A gap with no content behind it goes to course tuning, not into a path that pretends to cover it.
Session Coaching
Nobody is asked anything hereAgents move each learner's difficulty on their answers inside the session, coach the step a learner is stuck on, and refuse to hand over the answer to anything graded. Nothing waits on a person here. A learner who stalls gets the step explained in the session, not a ticket answered next week.
Submission Grading
Nobody is asked anything hereAgents grade every submission against its rubric the day it lands, write feedback that names the step that went wrong, and hold every grade that sits on an edge a person should call. Nothing waits on a person here. A held grade reaches the instructor's desk at station 5; everything clear of the edges releases with its feedback the same day.
Instructor Briefing
A person answers hereAgents score every learner's week against their own baseline, put the sliding ones on one desk with the held placements and held grades beside them, and draft the outreach in the instructor's voice. This is where a person works. Everything else exists so that six learners, not two hundred and thirty-eight, reach this desk.
Course Tuning
A person answers hereAgents trace every stall, quit and repeated tutor question back to the step that caused it, propose the content change that would have removed it, and propose nothing where one learner stalled once. A person accepts or rejects each change. What the course teaches, and in what order, is not something an agent changes quietly.
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:
When a placement contradicts itself and the system averages it anyway, the learner starts a course at a level nobody actually measured. They spend weeks bored or drowning before anyone notices the placement was fiction.
When a path cannot finish inside the learner's window and the system stretches it instead of saying so, the learner finds out at the deadline, not at enrollment. By then the time to renegotiate is gone.
When the tutor hands over the answer to a graded question, the grade stops meaning the learner can do the work. Every certificate downstream inherits that doubt.
When a score one point off a pass line gets rounded instead of held, a machine changes someone's transcript and no person ever sees the evidence. Matching answers slip through the same gap, so copied work passes ungraded.
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.
placementlearners/<id>/placement.jsonhold-contradictory-placementlearners/*/placement.json, the question bank and skill list read-onlypath-builderlearners/<id>/path.jsongap-without-contentlearners/*/path.json, content/modules/** read-onlysession-floorlearners/<id>/sessions/<date>.jsongraded-answer-requestedlearners/*/sessions/**, content/** read-onlygraderlearners/<id>/grades/<submission-id>.jsonhold-borderline-gradelearners/*/grades/**, rubrics/** read-onlyintervention-deskdesk/<week>/<learner-id>.mdawait-instructor-calldesk/**, everything else read-onlycourse-tuningchanges/proposed/<change-id>.json, reports/course-<date>.jsonaccept-course-changechanges/proposed/**, reports/**, the history read-onlyStack
Every system this agent reads or writes.