# Agentic Ad Operations: Research -> Brief -> Make -> Ship as One Loop
> An agentic loop isn't AI doing the work -- it's four stages sharing one state. Here's where the handoffs leak, what actually got faster in our own runs, and the two generations that scored 65 and 45 and never shipped.
- **Author**: Marxx
- **Published**: 2026-10-12
- **Category**: Product
- **URL**: https://marxx.ai/posts/agentic-ad-operations
---
Most ad teams run four processes that pretend to be one. Research happens in a deck. The brief happens in a doc. Production happens in Figma and a drive folder. Shipping happens in Ads Manager. Each step produces a file, and the next step reads about a third of it.
An agentic loop is not "AI does the work." It is **the four stages sharing one state**, so nothing has to be re-explained at a boundary.
Here's what that actually looks like, including where ours still breaks.
## The loop
The arrows aren't the interesting part -- every team already has arrows. The strip along the bottom is the argument. When the decode, the brand rules and the review criteria live in one place, the shipping stage can still see what the research stage found. When they live in four tools, it can't, and the loop is just a flowchart.
## Where the work is actually lost
Nobody loses a week *inside* a stage. They lose it at the boundaries.
Look at the first row especially. In one decoded account, ads built on "it's expensive" averaged 12.0 days live while ads built on "the battery dies" averaged 21.8. That difference was decided in the handoff from research to brief -- and almost every brief template in circulation has no field for it, so it falls on the floor every single time.
The fourth row is the quiet one. If nothing records *which angle an ad was testing*, then in a month you have performance data and no hypothesis to attach it to. You learn that ad 4 beat ad 7 and nothing about why. The loop never closes, which means it was never a loop.
## What actually got faster, and what didn't
This is the part the category oversells, so here are our own numbers.
One generated ad asset: **6 credits, under a minute.** Then a quality check against the brief and brand rules. It **scored 65 out of 100 and failed** on four issues -- and the fault was ours: the generation prompt specified green bars and a pink glow while the brief declared a periwinkle-only palette.
We fixed that and regenerated. Version two scored **45** -- worse. Fixing the brand problem broke the communication problem; brief adherence fell from 50 to 30.
Two attempts, 12 credits, about four minutes, neither shipped.
That is the honest shape of agentic ad ops. **Production stopped being the bottleneck and judgment became it.** The loop doesn't remove the human -- it moves them from making things to deciding which things are worth shipping, and that job gets harder as volume rises, not easier.
The value is real but it's narrower than the pitch: being wrong now costs four minutes instead of two weeks. That changes how many angles you can afford to try. It does not mean the first output is right.
## It also just breaks sometimes
Mid-run, our connector authentication dropped and the third generation never happened. Nothing recovered it automatically; someone had to notice.
Any honest account of agentic operations includes this. Loops that touch four systems inherit the failure modes of all four. Build a visible failure state -- a run that stops should be loud, not silently absent -- or you will ship nothing on a Friday and find out on Monday.
## Where to put the human
| Stage | Run it agentically | Keep a human |
|---|---|---|
| Research | Yes -- decoding a library by hand is wasted time | Choosing which question to ask |
| Brief | Yes -- if the research state is attached | Approving the problem and the objection |
| Make | Yes -- this is where the speed is | Nothing, until review |
| Review | Partly -- score against explicit rules | The call on what ships |
| Ship | Yes -- publishing is mechanical | Budget and audience |
The pattern: **automate the production, keep the judgment, and write the rules down so the review step has something objective to score against.** Most failed agentic rollouts skip that last clause and then wonder why review turned into an opinion contest.
## Publishing is the easy part
Once an ad is approved, pushing a campaign, ad set and ad into Meta is mechanical and should be automatic -- it's the stage with the least judgment in it and the most clicking.
Our canvas runs the chain in one place: research, write, image, video, audio, talent, publish. One brief in, several finished variants out, with the decode and brand rules still attached at the end.
## FAQ
**Is this just automation with a new name?** No. Automation runs fixed steps. An agentic loop chooses steps based on what the previous stage found -- which only works if the state carries.
**Do I need MCP for this?** It helps, because it lets a model query your ad library and your brand rules directly rather than being pasted a summary. The principle holds without it; the fidelity drops.
**What's the realistic time saving?** On production, large -- minutes instead of days. On the full loop, smaller, because review is now the constraint. Expect to be wrong faster, not right immediately.
**Where do most teams go wrong?** Automating the making before writing down the rules. Then nothing can be scored, and the review stage quietly becomes the slowest part.
**How many variants should a loop produce?** As many as you can judge properly. Twelve generated assets and no review capacity is twelve pieces of unshipped work.
**Can it publish without a human?** Technically yes. We don't recommend it, and neither does anyone who has watched an unreviewed batch go live.
**Trial the loop end to end: a rival ad goes in, your ad goes live, in one run.** [Book a demo](https://marxx.ai/book-a-demo) -- bring one competitor and one product.
---
- [More Product articles](https://marxx.ai/posts/category/product)
- [All articles](https://marxx.ai/posts)
