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 Unit Inspection That Pulls Every Flagged Unit Before Packing
Every unit the line makes, from the moment it passes the cameras to the moment it ships, gets reworked, or gets scrapped. The agents film every unit, call every defect they can stand behind, pull every flagged unit off the line before packing, and hand a person only the units where the frames do not settle it. A person used to check one tray in ten; now every unit is looked at, and the person's eyes go where the cameras are unsure. Six stations, run as a loop: a threshold accepted at station 6 tonight is the line station 2 calls against tomorrow, and a camera flagged at station 1 is serviced before tomorrow's first unit.
What Was Broken
What We Built
Six stations and 13 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Line Capture
Nobody is asked anything hereAgents film every unit from both cameras as it passes, match each frame to its belt position, retake a bad frame while the unit is still in view, and flag a camera whose frames are going soft. Nothing waits on a person here. A camera going soft is flagged to maintenance before its frames go bad.
Defect Calling
Nobody is asked anything hereAgents find each unit in its frames, outline every mark, gap and misplaced part, name each one from the plant's own defect catalogue, and mark the unit unsure when the frames do not carry the call. Nothing waits on a person here. An unsure call goes to the hold bench with its frames, never to a guess.
Unit Tracking
Nobody is asked anything hereAgents follow each unit from camera to camera on its belt position, join its frames and calls into one record, and watch the shift for repeats no single unit shows. Nothing waits on a person here. A repeat that clears the rule becomes an alert at the next station, with its units attached.
Divert & Alert
Nobody is asked anything hereAgents route every called, unsure or disputed unit into the hold lane before packing, send the alerts the rules already name to the people the rules already name, and keep one problem to one alert. Nothing waits on a person here. The unit reaches the lane and the alert reaches a phone, and a line stop is handed to a person, never made.
The Hold Bench
A person answers hereAgents build the case for every held unit: the deciding frame with the mark on it, the call, the three closest past decisions, and a recommendation where the frames support one, left blank where they do not. This is where a person works. Scrap, rework and release are a person's words, and the bench never empties itself.
Shift Close
A person answers hereAgents count what the line made, what the lane held and what the bench decided, draw the heatmap of where defects landed, and propose the threshold changes today's releases ask for, with what each would wrongly pull written beside what it saves. A person accepts or rejects each change. Where the sureness line sits is the plant's word for what ships, and an agent never moves it 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:
A lens picks up dust and the frames stop being judgeable. Without the hold rule, the scorer treats the blur as a clean unit and ships everything the camera cannot see.
The model is 60 percent sure a weld is bad, and the system rounds that up to a pass. The hardest defects sit in that middle band, so rounding hides exactly the units a person needed to see.
An agent sees a defect run and halts the line on its own authority. One false alarm costs an hour of output, and after two of them the crew starts ignoring every alert the system sends.
A unit comes off the lane and nobody writes down who cleared it or why. Six weeks later a customer returns it, and the plant cannot show whether it was rechecked, reworked, or waved through.
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.
line-captureframes/<unit-id>/camera-out-of-serviceframes/**, the camera feeds read-onlydefect-callingunits/<unit-id>/verdict.jsonunsure-callunits/**, frames/** read-onlyunit-trackingunits/<unit-id>/trail.jsonunit-lost-between-camerasunits/**, line/history/** read-onlydivert-alertlane/<unit-id>.json, alerts/<alert-id>.jsonline-stop-recommendedlane/**, alerts/**, units/** read-onlybench-casebench/<unit-id>/case.mdawait-your-decisionbench/**, everything else read-onlyshift-closethresholds/proposed/<change-id>.json, reports/shift-<date>.jsonaccept-threshold-changethresholds/proposed/**, reports/**, the history read-onlyStack
Every system this agent reads or writes.