Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 4, 5 and 6 need a person. Step 4 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 Manufacturing Automation Company Built Part Inspection That Catches the Same Defect Coming Back
Every part that comes off the line, from the moment it passes the cameras to the moment it lands on the pallet, in the reject bin or on the hold tray. The agents photograph every part, grade every surface, place each one where its verdict says, watch the shift for a defect coming back part after part, and hand a person only the parts nobody can call, a line stopped on a defect run, and the rules for what counts as a defect. Six stations, run as a loop: what the tray corrects today is what the grader checks against tomorrow.
What Was Broken
What We Built
Six stations and 12 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Part Capture
Nobody is asked anything hereAgents photograph every part from four angles as it passes the tunnel, match each frame to its part off the belt encoder, and re-shoot any frame too blurred or too dark to grade. Nothing waits on a person here. A part the cameras never got clean goes to the tray marked not seen, never graded off a bad picture.
Defect Grading
Nobody is asked anything hereAgents find the part in its frames, name which product variant it is, check every surface against the defect catalogue, and call each part pass or reject with the frames behind the call. Nothing waits on a person here. A part the grader cannot call cleanly goes to the tray with what it saw, instead of a guess going to the pallet or the bin.
Part Sorting
Nobody is asked anything hereAgents compute the pick point from each part's position and turn, guide the arm to place passed parts on the pallet, rejects in the bin and tray parts on the tray, and count all three against the verdicts every hour. Nothing waits on a person here. A count that disagrees with the verdicts stops the arm, and the stop reaches the line lead through the Line Watch.
Tray Review
A person answers hereAgents build the case for each tray part: its frames, the three most similar parts already decided, the nearest catalogue defect and a drafted recommendation. A person decides keep, scrap, or back through the tunnel. This is where a person works. Everything else exists so that five parts, not two thousand, reach this tray.
Line Watch
A person answers hereAgents count every reject by defect, spot and press across the running shift, send a forming run to the line lead with the parts behind it, and stop the feed when the same defect keeps coming off the same press. A stopped feed waits for a person. Agents can stop a line; only the line lead restarts one, after the press has been looked at.
Shift Close
A person answers hereAgents reconcile every part that entered the tunnel against the pallet, the bin or the tray, draw the shift's rejects on an outline of the part, and propose the grading changes today's tray corrections point at. A person accepts or rejects each change. What counts as a defect, and when a run stops the line, 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:
The grader scores whatever frame it gets, sharp or not. A defect hidden by blur ships as a good part.
The grader picks a side on every part, even the ones it cannot read. A wrong guess sends a bad part to the pallet or a good part to the bin.
The count says something is off and the arm keeps placing. Every part placed after that mixes graded parts into wrong bins, and the sort has to be redone by hand.
The feed stops on a defect run and the agent restarts it before anyone checks the press. The same defect runs again at full speed into the same bins.
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.
shoot-and-checkline/frames/<part-id>/part-not-seenline/frames/**, the camera feed read-onlyfind-and-gradeline/verdicts/<part-id>.jsonno-clean-callline/verdicts/**, line/frames/** read-onlypick-and-confirmline/placed/<part-id>.jsoncount-mismatchline/placed/**, the arm controller, line/verdicts/** read-onlytray-casetray/decisions/<part-id>.jsonawait-tray-decisiontray/**, line/** read-onlywatch-the-linealerts/<alert-id>.jsonfeed-stopalerts/**, the feed's stop line, line/verdicts/** read-onlyclose-the-shiftrules/proposed/<change-id>.json, reports/shift-<date>.jsonaccept-grading-changerules/proposed/**, reports/**, the shift history read-onlyStack
Every system this agent reads or writes.