Marketing OS
The centre node on page 1 of ../roadmap_ref_extracted.md,
and deliverable ④ on page 5: "Marketing OS — Architecture".
Two things live here, and they are two halves of one idea:
| What it is | Status | |
|---|---|---|
| Intelligence Layer | One shared store of what the business knows — rules, brand, keywords, personas, competitors, choices, performance | ⬜ Proposed (one instance ships) |
| Agents | The reasoning steps that read from it and act | ✅ 3 built · ⬜ 7 proposed |
The layer is the authority. The agents reason over it and must show their working.
That is not aspiration — the compliance rulebook already works this way, and throws away any citation the model invents.
Start here
| If you want | Read |
|---|---|
| The architecture idea, in plain language | Intelligence Layer — overview |
| How to build it | Intelligence Layer — technical |
| Every agent, built and proposed, with sequencing | Agent roster |
| One specific agent | Agents index |
| How this maps to ATL / BTL / Digital / Program / Analytics | Workflows |
Why they are documented together
Each agent today assembles its own context — the blog generator separately fetches brand voice, compliance rules, a competitor brief and SEO keywords, then prunes each to fit its own budget. That is workable at three agents and becomes the main source of bugs at nine: one fact changes, nine places it might not take effect.
The layer inverts that. Agents stop gathering and start asking.
Which means the agent roster and the layer cannot be planned separately — three of the seven proposed agents get materially cheaper if the layer lands first, and one of them (Learning) exists mainly to write back into it.
The one architectural decision
"The data becomes the LLM" splits into two very different systems:
- Retrieval — agents look facts up, cite them, and the store stays editable.
- Fine-tuning — facts are baked into model weights.
For facts and rules under RBI scrutiny, retrieval. An answer that cannot cite its source is not defensible in review, and a wrong fact in weights survives until retraining.
Fine-tuning stays open for style — house voice, script cadence — where there is nothing to cite.
Full reasoning: intelligence-layer/overview.md.
Structure
marketing-os/
index.md ← this file
intelligence-layer/
overview.md ← plain-English
technical.md ← proposed design + migration path
agents/
index.md ← the three built agents
roster.md ← all nine, tiered and sequenced
<name>-agent/ ← README · overview · technical
proposed/<name>-agent/ ← README (spec stubs)