MIZ OKI 3.5 Cloud Run Production Plan — Agentic Marketing Control Plane
Executive intent
This plan translates the latest agent-orchestrated marketing patterns into a concrete production path for MIZ OKI 3.5 on Cloud Run. The target outcome is not more disconnected automation. It is a governed, semantic, multi-agent operating layer that can observe demand, coordinate cross-channel actions, enforce policy, and learn from outcomes in near real time.
Target production architecture
1. Semantic marketing control plane
Deploy a shared control plane on top of the existing event graph pipeline so every decisioning agent reads and writes to the same marketing state substrate.
Control-plane domains - Customer graph: identity clusters, journey stage, behavioral context, consent flags. - Campaign graph: campaign hierarchy, budget state, attribution state, pacing, experiments. - Content graph: creative variants, asset lineage, message frames, offer taxonomy. - Constraint graph: budget policies, brand rules, legal constraints, rollout thresholds. - Memory layers: episodic outcomes, semantic relationships, procedural playbooks.
Cloud Run fit
- Keep marketing-event-graph-service as the public orchestration facade.
- Use Boss MCP as the policy-aware tool-routing layer.
- Persist semantic state to the KG system of record already owned by cell03/cell30.
2. Multi-agent coordination model
Run specialized roles with explicit contracts rather than one generalized planner.
| Agent | Responsibility | Primary runtime dependency |
|---|---|---|
| Strategy agent | Goal decomposition, scenario planning | Boss orchestration + replay/OPE guardrails |
| Audience agent | Identity-informed segmentation and prioritization | cell03, cell11, cell30 |
| Creative agent | Variant selection and message framing | content graph + policy constraints |
| Channel agents | Search, social, email, CRM, web execution | channel adapters + budget policies |
| Analytics agent | Performance, attribution, incrementality | cell15, causal/OPE services |
| Optimization agent | Budget reallocation and experimentation | operational brain |
| Governance agents | Compliance, brand, risk, explainability | policy layer + audit ledger |
3. OODA/SRPVDAL-aligned runtime loop
- Observe: ingest raw channel, web, and CRM events.
- Orient: refresh semantic control-plane state and customer/campaign memory.
- Decide: score intent, simulate budget moves, and select next-best actions.
- Act: route decisions through cross-channel adapters.
- Learn: write outcome evidence back into episodic/procedural memory.
Recommended implementation sequence
Phase 1 — Production control-plane substrate
Objective: make semantic state first-class.
Implement - Extend the marketing event graph flow with a control-plane snapshot API. - Normalize channel budget state into one canonical mix object. - Capture governance requirements and memory-layer summaries next to journey outputs. - Expose a control-plane payload for downstream Boss, UI, and policy services.
Success criteria - One API returns customer graph, campaign graph, content graph, constraint graph, and memory metadata. - Boss can inspect state without bespoke joins across services. - UI can render orchestration posture from a single response.
Phase 2 — Multi-agent orchestration contract
Objective: make specialized agents explicit and inspectable.
Implement - Add an orchestration contract describing strategy, audience, creative, analytics, optimization, channel, and governance roles. - Keep channels abstracted behind execution interfaces rather than direct agent-specific API knowledge. - Return a deterministic execution plan and decision-loop trace for every orchestration run.
Success criteria - Every run contains an auditable agent topology. - Budget and activation recommendations remain policy bounded. - Governance agents are embedded in the same response as execution actions.
Phase 3 — Cross-channel execution abstraction
Objective: turn channels into execution endpoints.
Implement next
- Formalize adapters for email, search, social, crm, and web under one orchestration interface.
- Attach channel-specific retry/fallback metadata at the adapter layer, not in agent logic.
- Record execution lineage for every action request.
Phase 4 — Memory and learning infrastructure
Objective: move from one-off decisions to compounding performance.
Implement next - Persist campaign outcome summaries as episodic memory. - Persist semantic entity and relationship updates in KG-backed memory. - Promote high-performing decision patterns into procedural playbooks. - Feed learning artifacts into replay/OPE and policy review loops before broad rollout.
Deployment mapping for MIZ OKI 3.5
| Concern | Current production anchor | Recommended extension |
|---|---|---|
| Event ingestion | cell02 |
keep as raw extraction edge |
| Event normalization + KG writes | cell03 |
remain canonical semantic write path |
| Identity resolution | cell03 + cell30 |
enrich with confidence-based control-plane inputs |
| Journey state | cell11 |
expose journey-stage summaries to control plane |
| Attribution | cell15 |
feed optimization + governance evidence |
| LTV scoring | cell20 |
feed prioritization and budget logic |
| Orchestration facade | marketing-event-graph-service |
add control-plane + agentic orchestration endpoints |
| Global tool routing | Boss MCP | register new MCP tools and keep registry authoritative |
Governance requirements
All production runs should remain: - auditable via lineage and decision traces, - explainable via explicit reasoning summaries and named governance agents, - reversible via bounded budget shifts and rollout controls, - policy-safe via embedded compliance, brand, and risk checks.
What is already integrated in this repository
The current implementation now adds:
- a semantic control-plane builder to the marketing event graph toolchain,
- a full agentic orchestration wrapper around the existing end-to-end pipeline,
- service endpoints for /control-plane and /agentic-orchestration,
- explicit governance-agent and decision-loop outputs,
- UI-level overview content for the new marketing control tower architecture.
Immediate next rollout steps
- Route Boss discovery toward the new orchestration endpoint for production-safe trials.
- Connect live channel adapters behind the orchestration contract.
- Add replay/OPE gate checks before autonomous budget changes exceed tenant thresholds.
- Wire control-plane snapshots into the command-center dashboard for operator review.
- Add persistent storage for episodic and procedural memory artifacts.