A week inside the agent pipeline that runs our own marketing | APX Digital

We run our own marketing on agents: a week inside the pipeline

Every agency tells you they use AI internally. Here is the receipts version: our social pipeline has been running agents since 3 September 2026, and after its first day it had published exactly nothing. That was deliberate. This post walks through the actual files on our disk: the queue, the planner, the render, the QA gate, and the two kill-switches that held the output hostage, because we'd rather show you the machinery than sell you the idea of it.

What the pipeline is

One repo, four scripts, one playbook. Every day, an agent orchestrator:

1. asks a planner which post is due,
2. writes or restyles the creative,
3. renders it to PNG in a headless browser,
4. looks at the render with a vision model and fails its own work until it's clean,
5. stages everything in a dated drafts folder,
6. tries to publish, and for now correctly refuses.

The whole thing is under a hundred lines across four scripts, plus a written playbook (672 words of instructions the agent must follow exactly, including the line "never post invented claims"). Everything below is verifiable from files on disk, measured 3 September 2026.

The planner: rotation, not vibes

The queue holds 14 entries: each a service, an angle, a format (single, portrait, carousel, reel) and a designed background. The planner script is deterministic: it keeps a cursor, hands out the next entry each day, and is idempotent per date: run it four times on the same day and you get the same assignment without advancing. It also alternates backgrounds across a fixed cycle (white, black, acid green) and never assigns the same polarity twice running. When the cursor wraps past entry 14, a cycle counter ticks up, and cycle 2 forces the agent to rewrite every caption from a different angle and restyle any reused art, so the grid doesn't repeat itself.

No LLM chooses what runs today. A script does. That is the first lesson of running agents on real work: decisions you can make with a loop should never be made with a model.

Staging before anything else

Each day produces an exact folder: daily/drafts/YYYY-MM-DD/ containing a manifest (what ran, what was reused, QA result, per-platform publish flags all starting `false`), platform-native copy in separate files, the final PNGs, and a five-line brief addressed to the human. Day one's brief says, verbatim: "Publish: STAGED, no Meta page token or xurl auth on this machine yet. Needs Jamie: add FB page ID to state/config.json..." The agent knows its own blocker and escalates it in writing instead of improvising around it. That behaviour is not magic; it's a line in the playbook.

The render and the QA gate

The creative is HTML in a strict design system: Archivo 900 uppercase, Space Mono rails, hairline rules, exactly one acid-green element per composition. A 10-line script feeds it to headless Chrome at an exact pixel size, with an eight-second virtual-time budget so webfonts finish loading before the screenshot. Then the QA gate: a vision model reads the finished PNG against a checklist (fonts loaded, no clipping, one accent only, no typos), and the render is redone until it passes. Two consecutive failures and the pipeline falls back to a library asset and says so in the brief. No silent degradation.

The kill-switches nobody wired up on purpose

This is the part most "AI marketing" demos skip. Publishing requires two things the agent cannot install for itself:

• push-meta.sh (25 lines) posts to the Facebook page via the Graph API. It exits blocked if the page ID field in config is empty (it is empty) or if a long-lived page token is missing from the macOS Keychain, which the script fetches with security find-generic-password. The agent never sees the token; it only gets to run a script that uses it.
• push-x.sh posts to X only if a separate CLI has valid stored auth. It doesn't.

So the pipeline can write, render and QA a full week of campaigns and put none of them online until the owner types a page ID into a file and pastes a token into the Keychain. We built it that way. The first history line in state/history.jsonl reads `"published": "none"`, a machine-readable record that nothing escaped. When Jamie adds credentials, publishing turns on one platform at a time, and every prior post stays reviewable in drafts first.

If you're evaluating an agency that runs agents on your accounts, ask to see their own kill-switch. Ours is on disk.

What a week of this costs, honestly

Day one produced a five-slide carousel plus fallback single, all from the existing library with fresh captions and no new art, because the planner said the angle was already covered. The agent's working day is minutes of compute; the human's involvement is the brief and, when gates fail, the fix list. The output floor is one post a day, every day, at a consistency the brand system enforces, and a publish rate of zero until a human says otherwise. Both halves of that sentence matter.

The definitional groundwork (what actually counts as an agentic system) is in What an agentic system is (and is not). The buyer's checklist, distilled from how this pipeline is fenced in, is What to demand before an agent touches your inbox.

---

Want this pointed at your business? APX builds agentic systems, starting with chatbots (Basic £397 · Advanced £797 · Enterprise £1,497 one-time, support from £50/mo, apxdigital.io/ai-chatbots); pipeline work is quoted on scope. Info@apxdigital.io · 07753 226214 · Birmingham / Coventry / West Midlands.