Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 1, 3 and 6 need a person. Steps 1 and 3 are waiting for approval right now. Click either 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 HR Services Company Built Content Publishing Where Every Draft Carries Its Sources
Keeps the publishing calendar fed without letting the brand voice thin, and never lets a claim it could not source go out under the company's name. Written for a marketing team that publishes more than it can hand-check: every draft carries its sources, every number on the dashboard carries its trace, and the person's work shrinks to picking topics, clearing drafts and signing the quarterly review. Six stations, run as a loop: what station 6 learns re-aims what station 1 briefs.
What Was Broken
What We Built
Six stations and 14 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Topic Selection
A person answers hereAgents track what the site ranks for, read what outranks it, and draft a brief for each topic worth a calendar slot. The person picks which briefs earn a slot. The calendar is a promise the team makes, so an agent does not make it.
Draft Writing
Nobody is asked anything hereAgents write each committed brief into a full draft, pull it into the house voice, and validate it against the article schema before it can move. Nothing waits on a person here. A draft reaches somebody at station 3, and only after every gap in it is closed.
Editorial Review
A person answers hereAgents open every source the draft cites, route the legal-adjacent passages to the compliance reader, and put each waiting draft on one card the editor can decide from. This is where a person works on every piece. The byline is a person's word, and an agent cannot give it.
Channel Publishing
Nobody is asked anything hereAgents slot each cleared draft into the calendar, cut the newsletter and social versions from the live page, and check every version against the page it came from before it goes out. A person signs nothing here. The signature already happened at station 3, and a failed live check pulls a version back on its own.
Performance Tracking
Nobody is asked anything hereAgents pull rankings, traffic and conversions for every live piece, trace each lead back to the pages that carried it, and test title and call-to-action variants where a page underperforms its position. Nothing waits on a person here. The week's numbers land in Monday's digest, and only traced numbers land at all.
Library Review
A person answers hereAgents sort the whole library into keep, refresh and retire with the evidence for each call, queue the refresh briefs back into station 1, and draft the quarter's report. A person signs the quarter. Retiring a page and redirecting its address is hard to reverse once other sites link to it, so that call stays human.
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 agent cannot verify a term's volume, so it carries a number anyway. Every topic decision downstream now rests on demand nobody measured.
The brief calls for a stat and no source holds it, so the draft makes one up. The invented fact reads exactly like a real one. It ships under the company's name.
The reviewer cannot open the source and passes the claim anyway. The first person to actually check it is a customer.
A lead with a broken path gets handed to whatever page sits closest. The report says that page works when nobody knows if it does, and next quarter's budget follows the report.
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.
topic-scoutpipeline/briefs/<brief-id>.jsoncalendar-commitpipeline/**, library/** read-onlydraft-writerpipeline/drafts/<draft-id>.mdunsupported-claimpipeline/**, voice/** read-onlyreview-briefpipeline/review/<draft-id>.jsonawait-editor-decisionpipeline/review/**, everything else read-onlypublish-runpublished/<slug>/manifest.jsonlive-check-failedpublished/**, pipeline/review/** read-onlyperf-readeranalytics/weekly/<week>.jsontest-inconclusiveanalytics/**, published/** read-onlylibrary-reviewlibrary/review-<quarter>.jsonsign-quarter-reviewlibrary/**, analytics/** read-onlyStack
Every system this agent reads or writes.