AI Agents · AGENTIC-MARKETING · Building

Marketing Agent.

Goal in plain English → audience, creative, channel mix, send time, measurement — autonomously, with humans only on guardrails.

Building SELL L4 Powered by → UCP

Agent topology

11 agents

Meta-Orchestrator + 9 Domain + Compliance

Cycle target

Hours

in-development · vs weeks today (11 teams · 6 handoffs)

Decisioning

1:1

real-time per-customer · not batch segments

Compliance gate

Always-on

consent · frequency · sensitive-category

01 · The job to be done

The pain

Today: 11 teams, 6 manual handoffs per campaign, weeks-long cycles. Marketing intelligence lives in PowerPoints, not systems. Each new campaign starts from scratch — no compounding learning. By the time performance data lands, the moment to course-correct has passed.

What the agent does

A marketing leader states intent + budget + dates + objective. The Meta-Orchestrator decomposes into sub-tasks. 9 Domain Agents collaborate on research, audience intelligence, planning, content, channel orchestration, experimentation, measurement, opportunity detection, and dashboard intelligence. A Compliance Agent enforces governance at every step. Outcome signals from each campaign feed back into the signal layer — knowledge compounds across runs.

What humans still do

Marketing leadership defines brand objectives, budget, brand rules, sensitive-category policy, approval thresholds. The agent runs the campaign loop. Humans intervene on guardrail violations and on strategic intent.

02 · How it works · the closed loop

Inputs (blue) feed Meta-Orchestrator, which fans out to 9 Domain Agents. Compliance + Approval gates (yellow) sit before activation. Outcome signals close the loop back to the agents.

Marketer states intent

Goal in plain English ("Diwali sale for fashion · 30-50yr customers in Mumbai"), budget, dates, objective. No segmentation, no creative briefs.

Meta-Orchestrator decomposes

Parses intent · creates a workflow DAG · routes parallelisable tasks (Research, Audience) to specialists, sequencing dependents.

9 Domain Agents collaborate

Research mines prior campaigns. Audience builds segments + lookalikes. Planning drafts the brief. Content generates creatives. Channel selects mix + send-time. Experimentation wires A/B + holdouts. Measurement attributes outcomes. Opportunity detects anomalies. Dashboard surfaces summaries.

Compliance Agent gates

Cross-cuts every action — consent enforcement, frequency caps, sensitive-category rules, brand-voice guardrails. Blocks any action that violates policy.

Approval gate

Marketer reviews audience, creative, channel plan, spend forecast. Approves; activation proceeds. Rejects with notes; agents iterate.

Activation

Campaign launches across channels — SMS, WhatsApp, RCS, push, email, paid, in-app. Each personalised to person + household + context.

Outcome → signals

Performance signals (clicks, conversions, lift, churn saves, revenue) write back into the signal layer. Next campaign starts from a smarter base — knowledge compounds.

03 · Underlying data

Data source	Classification	Key entities
UCP signals	Mixed	Identity · behavioural · transactional · contextual · household · preferences · AI-derived traits
Channel platforms	Machine-Readable	CleverTap · Google Ads · Meta · WhatsApp Business · RCS gateway
Databricks	Machine-Created	Customer transaction tables · computed CLV · segment definitions
Real-time context	Machine-Created	Store proximity · inventory availability · festival windows · device
Outcome history	Machine-Created	Prior campaign signals · A/B · holdouts · bandit arms

Refresh cadence: Identity stitching real-time; signals continuous; outcome attribution near-real-time post-send.

KPIs moved:

Cycle time (weeks → hours)Campaign ROI / lift on holdoutsSuppression and frequency-cap complianceCompounding selection accuracy across runs

04 · Design patterns used

Multi-Agent Orchestration

Meta-Orchestrator coordinates 9 Domain Agents + 1 Compliance Agent · peer collaboration with explicit handoffs

Tool Calling

Channel APIs (CleverTap, Google, Meta, WhatsApp, RCS), UCP signal queries, content generation, outcome attribution — all tool-mediated

Agent Memory

Per-agent Firestore memory · prior campaign decisions feed next run; episodic + procedural

Wide Research

Research + Audience agents explore many segment / lookalike candidates before Planning narrows

Observability

Every agent decision traced · token-economical artefact-passing keeps context lean

Evals

Holdouts and A/B + bandit experiments built into every campaign by default

05 · Evals · how we know it works

Per-campaign holdouts and A/B suites will measure incrementality at run time. Cross-campaign: backtest agent-decided audiences and creatives against historical campaign outcomes; compare lift versus the manual baseline. Compliance Agent's gate-pass rate is its own audit metric. Current state: framework documented; first end-to-end pilot campaign in setup. No measured numbers yet.

Gates and thresholds.

Lift on holdouts ≥ baseline manual campaign lift
Compliance violations: zero pre-activation; flagged within 1 hr post-send
Frequency-cap and consent enforcement: 100% adherence