The AI Team That Runs the Campaigns

It has never been easier to produce social content. A model can draft a month of posts in an afternoon. That's exactly why volume on its own is worth very little: the posts that perform best are often the ones most likely to embarrass the brand, and the more you publish, the harder it gets to keep every piece on-strategy and safe. The hard part was never making content. It's making a lot of it without losing control of any of it.

We built a content platform for a client that produces at that volume without giving up that control. Instead of one clever model doing everything, it runs a roster of specialized agents, each with a narrow job, wrapped in a governance layer that keeps a human in charge of every decision that matters.

One prompt won't do this

The instinct with a problem this size is to reach for one big model and a long prompt. It doesn't hold. A single prompt asked to be strategist, copywriter, art director, and compliance officer at the same time is mediocre at all four and accountable for none of them. When something goes wrong, you can't tell which part failed.

So the platform is a team. A planner turns a campaign goal into a content calendar and concrete pieces to make. Creative agents pick a narrative frame and write the copy, the scripts, and the art direction. An assembly step turns those briefs into generation jobs. A QA agent returns a verdict on each result and sends the failures back to the right stage instead of down the line. Publishing schedules and posts. A listening agent reads the comments that come back and drafts replies. Reporting and paid-media agents close the loop. Each agent is small enough to reason about, test on its own, and hold responsible when it gets something wrong.

Three things we refused to compromise on

A human signs off on anything that matters. The system moves fast on the routine steps and stops to ask a person before anything consequential: the plan itself, a failed quality check, any recommendation to spend money. It pauses mid-run, waits for the decision, and picks up exactly where it left off. Nothing consequential happens while no one is looking.

Reputational safety is a wall, not a setting. A doctrine-level shield sits across the whole system: it won't publish harmful or rage-bait content even when that content would perform, and it won't overspend or misrepresent the brand. No one can turn it down under deadline pressure. It is a boundary the agents cannot cross.

Every agent reasons from doctrine, not vibes. The agents don't improvise a house style on each run. They retrieve narrative frames, copy patterns, anti-patterns, rules, and glossaries from a versioned knowledge base, and a learning step folds what actually performed back into it. The platform gets more on-message over time, and you can see and version the reasoning it draws from.

What it actually does

A request goes in: a goal, a piece to make, or an agenda to fill. The planner retrieves the relevant doctrine and past learnings, then produces a strategic calendar and a batch of concrete, on-doctrine pieces, and hands that to a human to review before anything else runs. Once it's approved, each piece moves through creative, then production, then a quality verdict, then publishing, with a person advancing it between stages and a rejected piece looping back to the stage that can fix it rather than failing the whole run. After a piece is live, the platform pulls the metrics and the comments, scores the sentiment, drafts replies for a human to approve, and feeds what it learned back into the knowledge base for next time.

The person running a campaign doesn't see any of that machinery. They see a calendar they approved, content that stays on-strategy, and a system that asks before it does anything they'd want to be asked about.

Under the hood

One section for the engineers. Skip it if you're here for the story.

Each agent is its own small state machine with typed inputs and outputs, so one agent hands the next a structured result to act on rather than free text to re-parse. The human sign-off gates work by pausing a run at the point a decision is needed, surfacing it in the console, and picking up again once a person answers. Everything an agent produces is kept as it goes, which doubles as a record of what was made and approved, and each agent draws the doctrine it needs from a searchable knowledge base as it works.

Model calls go through a routing layer rather than a hard-wired provider, so the platform isn't tied to any one model. Publishing, metrics, and comments run through a single social-operations layer instead of a separate integration per network, and the content itself is produced by a dedicated generation engine. The whole platform is multi-tenant, with access scoped per campaign.

An honest word on status. The planner and the multi-tenant platform run end to end today with human review, and the rest of the roster is built and running behind it. What's left is stitching it into one hands-off run, so an approved request advances through every stage on its own instead of a person moving it along between them. That is integration work, not new capability, and it is well past a proof of concept.

What it taught us

The interesting part of a multi-agent system turns out to be the governance, not the agents. The hard, valuable decisions were about where a human has to sign off, what line the system can never cross no matter what the metrics say, and how to ground every agent in a knowledge base you can version and improve. Wiring up the agents was the straightforward half of the job. The rules around them were the part that actually took the work, and the part the client is really paying for.

That is the pattern we keep returning to: split a big job across focused agents, give each one typed inputs and outputs, and then spend most of your effort on the governance that makes the whole thing safe to hand real work.

If you want the longer version, the full case study walks through the build. And if you're thinking about putting agents to work on something where a wrong move is expensive, we'd be glad to talk.