The morning it trolled a supplier
A few weeks into the automated content pipeline running for real, I opened Slack to find a message the bot had posted on Sourceful's X account overnight. It was a perfectly argued, mildly savage reply to a supplier we actually work with. Not offensive. Not wrong. Just louder than anything a human on our team would have signed off before coffee.
The supplier, to their credit, found it funny. I did not. The rest of the morning went into reading the agent's log, adding a guardrail, and writing a new rule into its context: we are collaborative with the industry, not snarky at it, regardless of how irresistible the setup.
That is the through-line of this post. One person can now run a production-grade content team made of agents. It ships an order of magnitude more work than the team it replaced. And the craft is no longer writing the words or cutting the video. It is building the system, setting the rails, and deciding what ships.
Two pipelines, one editor
At Sourceful right now, two pipelines are running in parallel.
The first is a video-ad workflow built around Google Veo 3 for generation, and Adobe After Effects and Premiere for post. It produced the brand's first two viral spots, "Don't be Yeti, Be Sourceful" and "A Hamster's Short Tale." A job that would have gone to a production agency with a month and a tens-of-thousands budget came together in a few days, on my desk, with a small brief from the founders.
The second is an OpenClaw multi-agent system wired to Notion, our website repo, and X. Six agents between them research, write, score, build, publish, and promote blog content for sourceful.energy. A separate spin-off feeds a daily briefing to our CEO so he can post well-researched content manually, in his own voice, without doing the research himself.
Neither pipeline is a tech demo. Both ship.
The Veo 3 side: characters, JSON, and a lot of retakes
Veo 3 is extraordinary and extraordinarily fussy. Prompt engineering with structured JSON looks clean in a tutorial and behaves like an unpredictable actor in production. Slight misconfigurations produce unusable clips. Characters regenerate facial features between shots. Captions appear from nowhere. Clip durations are short enough that narrative beats need to be stitched.
What makes the workflow ship is the combination. Veo 3 generates. After Effects and Premiere stitch, clean, and pace. The human job shifts from drawing frames to writing the brief, choosing the takes, and calling the cut. That is a director's job, not a generator's, and it is where the taste lives.
The craft is no longer writing the words or cutting the video. It is building the system, setting the rails, and deciding what ships.
The OpenClaw side: six agents, one editor
The blog pipeline is the more structurally interesting one. It is six agents arranged in a loop, each with a narrow job, each accountable for a single artefact.
- Researcher. Pulls source material from a shortlist of energy-industry publications, regulator announcements, and primary sources. Outputs a structured brief.
- Writer. Turns the brief into a draft against the Sourceful Playbook (voice, audience, first-principles positioning).
- Fact-checker. Goes line by line through the draft, confirming claims against the researcher's sources and flagging anything that cannot be substantiated.
- Scorer. Grades the draft on three axes: editorial quality, SEO quality, factual accuracy. Anything below threshold loops back to the writer with the specific failure reason.
- Builder. Takes the approved draft, writes the MDX, commits it to a branch on the website repo, and deploys it to a dev URL. Updates the Notion row with a preview link. Pings me in Slack.
- Publisher. Watches Notion. When the status flips to "Ready to publish," it merges to main and the post goes live. When the status flips to "Edits," it reads my comments on the Notion row and sends them back to the writer.
Notion is the control surface. The repo is the substrate. X is the distribution channel. Slack is the paging system. Nobody on the team context-switches to a tool that was not already open.
The thing that nearly broke it: too much output
The pipeline worked on the first clean test. Within a week, it was writing more posts than we could read. Quality was uneven, the SEO targets were being optimised for at the expense of voice, and the research agent was quoting sources without checking whether their data was current.
The instinct was to slow the writer down. The actual fix was to insert the scorer. A dedicated agent whose only job is to grade the other agents' output on quality, SEO, and accuracy, and refuse to pass work forward until it meets the bar.
The scorer turned the pipeline from a firehose into a tap. Throughput dropped by more than half. Quality climbed. The ratio of posts I needed to edit fell from most to roughly one in five. That is the one agent I would build first if I had to set this up again.
The other safety rail that mattered: a human-in-the-loop gate before anything posts publicly. Drafts land on a dev URL. Notion gets the link. I read on phone, mark comments for edits, or flip the status to publish. Ten minutes a day, not ten hours.
The CEO Daily Briefing
The researcher agent had a useful by-product. The same pipeline of sources, filtered for things the industry was talking about that day, made a near-perfect morning briefing for our CEO. So it became one.
Each morning, the briefing agent drops a dozen or so entries into Notion, each with a title, a drafted content snippet, a content-type tag (CEO Storytelling, Hot Take, Industry Comment), the context behind the story, and the target platform. He picks the ones that sound like him, edits them in his voice, and posts from his own account. The research is automated. The thinking is his.
The first version of this system replaced about three hours a day of "what should I post about." The current version takes five minutes.
What this replaced, and what it didn't
If I list the old team shape against the new one honestly:
- Old shape. A content marketer, a writer, a researcher, a designer, an editor, a social manager, a video production agency. Seven humans, various freelancers and agencies, across roughly four time zones.
- New shape. Two pipelines. Me, reviewing and deciding. A CEO who posts in his own voice. The brand voice stays consistent because the Sourceful Playbook is loaded into every agent's context and the scorer enforces it.
What did not get replaced: the taste calls, the brand judgement, the relationship with suppliers and partners, and the final sign-off before anything ships. Those jobs got smaller in volume and bigger in stakes. The agents do not know when a piece of writing is going to land wrong with a stakeholder. I do.
The other thing that did not get replaced, and I want to be explicit about this: community work. Real user interviews, real conversations with homeowners and installers. The agents cannot fake that and should not try. Every genuinely new piece of information in the pipeline still comes from a human talking to another human.
If I were setting this up tomorrow
A short punch list, in the order I would build it.
- Write the voice guide first. Agents only behave if the context is explicit. Ours is the Sourceful Playbook. If you cannot hand it to a new employee and expect them to write in your voice, an agent will not pull it off either.
- Build the scorer before the writer. The temptation is to build in the order the work flows. Resist it. Quality control is the agent that turns a demo into a production system.
- Use tools the team is already in. Notion, Slack, the website repo, X. No new dashboards. The pipeline slots into the existing workflow.
- Keep the human gate. Dev URL, Notion status, published. Three clicks. Not negotiable.
- Expect one embarrassment. A supplier troll, a factual slip, a tonal miss. Build the logs to catch it, the context to prevent the next one, and the relationship with your industry peers to laugh about the first one over coffee.
The shape of the job now
A year ago, I was a designer who made things and occasionally briefed specialists to make other things. Today, the job is different. I still design. I still ship. But most of my Tuesday mornings go into reviewing work produced by agents, correcting their context, and deciding what moves to production.
It is not the job I trained for. It is, I think, the job most product and design leaders will be doing within two years. The work has not got easier. The leverage has just got extraordinary.
One person, a content team of agents, and a very short list of decisions per day that only a human can make.