Six-Step AI Workflow with Human Oversight
Each step shows the workspace, what the AI completed, and what still needs human review.
Steps 3, 5 and 6 need a person. Steps 3 and 5 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 Technology Company Built a Posting Loop with Human Sign-off Wherever the Brand Speaks
Keeps the brand talking on every network at the pace the platforms demand, and never lets an unsourced claim, an expired usage right, or an angry customer pass without a person. Written for a marketing team where one person used to write, post, watch and count everything by hand, and the feed went quiet whenever that person was busy. Six stations, run as a loop: what station 6 learns about the month is what station 1 plans the next one from.
What Was Broken
What We Built
Six stations and 15 subagents. Each subagent carries its own tasks and its own refusal.
How It Runs
Calendar Planning
Nobody is asked anything hereAgents watch what is moving on each network and draft the week's slots: topic, platform, format, day and time. Nothing waits on a person here. A slot only becomes a post, and every post stops at station 3 before it publishes.
Post Drafting
Nobody is asked anything hereAgents write one master draft per slot, cut it to each platform's length and register, and check every line against the voice guide before it moves. Nothing waits on a person here. Every variant stops at station 3 before it can publish.
Brand Sign-off
A person answers hereAgents assemble each post into one card a reviewer can decide from: every variant, every asset, the source behind every claim, the rights behind every clip. This is the station where a person works. Nothing publishes under the brand's name until somebody clears it.
Publishing
Nobody is asked anything hereAgents queue each cleared variant against its slot, publish it on its network at its time, and confirm the live copy matches the cleared copy. Nothing waits on a person here. A failed publish retries itself, and a failing network holds its own queue.
Community Engagement
A person answers hereAgents answer the routine questions from the approved answer bank, hide what the posted rules name, and escalate a thread that is heating up with its whole history attached. This is a station where a person works. An escalated thread gets no automated reply until somebody takes it.
Month Review
A person answers hereAgents pull the month's numbers from every network, reconcile them against what was actually published, trace what moved back to the posts that moved it, and propose the changes next month should run on. This is a station where a person works. No rule the agents run on changes until somebody accepts it.
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 brief arrives with no source and the writer fills the gap with a plausible fact. The post reads fine, and the claim behind it does not exist.
A photo license runs out and the post stays up. The rights holder finds it before you do.
A network's API starts failing and the scheduler keeps sending into it. Some posts go out twice, some are lost, and you cannot tell which.
A customer asks a question the playbook does not cover. The agent writes its best guess and answers as the brand.
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.
calendar-draftcalendar/<week>/slots.jsontopic-outside-brandcalendar/**, insights/** read-onlypost-draftposts/<post-id>/drafts/claim-without-sourceposts/**, assets/** read-only, calendar/** read-onlysignoff-prepposts/<post-id>/review-card.jsonawait-signoffposts/**, rights/** read-onlypublish-runpublished/<post-id>.jsonpublish-failed-twicepublished/**, posts/** read-onlyengage-runconversations/<thread-id>.jsonsentiment-escalationconversations/**, answerbank/** read-onlymonth-closeinsights/<month>-report.json, rules/proposed/<change-id>.jsonaccept-rule-changeinsights/**, rules/**, published/** read-only, conversations/** read-onlyStack
Every system this agent reads or writes.