Proprietary architecture

A multi-tenant, agent-first
orchestration platform.

46 proprietary sub-agents, each purpose-built for paid media. They reason, execute multi-step plans, interact with databases, manage campaigns, and delegate tasks to other sub-agents, all within strict security guardrails and human-in-the-loop approvals. Built from scratch, not assembled from third-party parts.

Request a demo

Two operating modes

Monitor mode watches. Act mode executes.
You choose when to switch.

Every sub-agent runs in one of two modes. Monitor connects to your platforms, observes everything, and surfaces what matters. Act takes the approved action. The human always decides when the system moves from watching to doing.

Agency principals

Scale your client base without scaling your team.

  • Deploy across every client account. One platform, 50 clients. The sub-agents monitor and optimise each one.
  • Scale from 10 clients to 50 without hiring. Your team focuses on strategy. Maaten handles the volume.
  • White-label it. Offer Maaten as your proprietary technology. Your brand, your client relationship.
  • Protect your margin. More clients, same headcount, better outcomes. The retainer model finally scales.

In-house performance leads

Replace six vendors. Answer the CMO in the meeting.

  • One platform replaces the stack. Retire Fivetran, Snowflake, Looker, Claravine, and your single-channel copilots.
  • Monitor mode holds your agency accountable. See what they optimise, what they miss, and what they could do better. Facts, not opinions.
  • Cross-channel attribution in real time. Not a Friday deck. Not a quarterly review. The answer is always current.
  • Act mode when you're ready. Start with visibility. Move to execution when the data gives you confidence.

Maaten-built, end to end

One closed loop.
Every layer Maaten-built.

Every layer designed, built, and maintained in-house. Each component feeds the others. Clean data makes smarter sub-agents. Smarter sub-agents produce better campaigns. Better campaigns generate richer data.

maaten-architecture
User describes business [A] Agentic Engine Builder sub-agent creates structure, hires specialist sub-agents [B] Naming Conventions Architect Sub-agents create campaigns with enforced naming, metadata, structure [C] Data Lake Raw data ingested, enriched, standardised, stored in Gold Tables [D] Intelligence Layer Sub-agents query data, detect anomalies, generate recommendations User reviews, approves, decides Sub-agents execute approved changes new data flows back into [C] Loop continues. Data compounds. Sub-agents get sharper.

Master-Node / Tenant architecture. The master node controls global configuration and deployment. Each client workspace is cloned from a master blueprint and runs independently.

Proprietary cognitive layer

Most platforms run one inference path.
Maaten built four.

Maaten dynamically selects from four distinct cognitive engines based on task complexity, urgency, and type. Each engine is proprietary. Each serves a different reasoning pattern.

Engine_ReAct

Reason and Act

The core autonomous loop. Maaten reasons about the problem, selects a tool, executes it, observes the result, and iterates until the goal is met. An Overseer model runs quality control on every cycle, catching drift and errors before they propagate.

Used for: campaign diagnosis, anomaly investigation, root cause analysis, creative performance review

Engine_PlanExecute

Plan, then Execute

For complex, long-term goals. Maaten writes a complete multi-step execution plan before taking any action, then completes tasks sequentially. Each step is validated against the plan before proceeding.

Used for: multi-platform campaign builds, cross-channel budget reallocation, quarterly reporting cycles

Engine_Swarm

Hierarchical Multi-Agent Routing

Maaten receives the prompt, decomposes it into sub-tasks, and delegates each to the specialist sub-agent best equipped for it. Results are composed back into a unified response. This is how Maaten coordinates multiple sub-agents on a single complex request.

Used for: cross-channel performance reviews, multi-agent campaign launches, requests spanning multiple domains

Engine_ZeroShot

Low-Latency Parallel Execution

No reasoning overhead. Simple, well-defined tasks execute in parallel at maximum speed. Used when the task is clear and the cost of deliberation exceeds the cost of execution.

Used for: metric lookups, status checks, format conversions, notification dispatches

Purpose-built tooling

~40 autonomous tools.
Written for paid media.

Sub-agents do not generate suggestions and stop. They execute. Maaten assigns specific tools to each sub-agent. No sub-agent has access to tools outside its scope. Every tool was written in-house.

Data and Memory

tool_rag
Extracts specific semantic chunks from uploaded knowledge bases via Qdrant vector search
tool_memory
Queries chronological context of the user and account across sessions
tool_query_database
Sub-agent writes and executes SQL/PgSQL queries against connected databases autonomously
tool_session_kv
Ephemeral scratchpad for multi-turn tasks. Persists within a session, cleared after
tool_save_knowledge
Self-teaching: the sub-agent saves solutions it discovers to future-proof its own performance

Action and Execution

tool_git_ops
Clone, commit, and push code to repositories
tool_manage_calendar
Generates .ics calendar invites on the fly
tool_send_email
Outbound SMTP communications with template support
tool_schedule
Writes delayed background tasks to the cron queue for future execution

Cognitive and Self-Reflective

tool_drift
Calculates cosine similarity of narrative focus to detect and correct sub-agent drift in real time
tool_prompt
Allows the sub-agent to securely rewrite its own system instructions when it learns a better approach
tool_diagnose
Sub-agent reads its own firewall and error logs to self-diagnose failures without human intervention

Web and File Operations

tool_crawl_website
Deep recursive spider mapping of domains
tool_generate_file
Converts raw data into downloadable PDFs or CSVs

Representative sample. The cognitive and self-reflective tools (tool_drift, tool_prompt, tool_diagnose, tool_save_knowledge) are proprietary to Maaten.

Proprietary memory system

Three memory layers.
Built into the orchestration layer.

Three proprietary memory systems run in parallel, each serving a different purpose.

Layer 1

Neural Memory

Immediate task context inside the model. What the sub-agent is working on right now. Scoped to the current session.

Layer 2

Semantic Memory

Long-term, vectorised memory stored in Qdrant. Account-level patterns, audience behaviour, creative performance signals, seasonal trends. Queryable by any sub-agent with access. Persists indefinitely.

Layer 3

SQL-Structured Memory

Persistent across all sessions. Queryable and joinable to the data lake. Where sub-agent learnings become part of the shared dataset. A bid strategy that overperforms gets reinforced. A creative angle that underperforms gets downweighted.

Retrieval-Augmented Generation (RAG): Sub-agents pull the right slice of the lake or knowledge base at query time, so answers are grounded in the customer's actual data rather than the model's training set. A proprietary WebGL interface renders the semantic memory as a navigable 3D structure for diagnostics.

The sub-agent catalogue

46 proprietary sub-agents.
All built for paid media.

Each sub-agent has scoped tool access, scoped data access, and governed actions. No hand-offs. No gaps. No third-party sub-agents.

Builder Sub-Agent. You describe your business. The Builder sub-agent creates your company structure, selects which specialist sub-agents to deploy, sets their parameters, and connects their data sources. Onboarding is itself agentic. No configuration forms. No setup project.

Monitoring and Detection ·11 sub-agents

Sub-agentStatus
Meta Monitoring AgentProduction
Google Monitoring AgentProduction
GA4 Monitoring AgentProduction
Pacing ControllerProduction
Optimization ScannerProduction
Enhanced Anomaly DetectorProduction
Consistency GuardianProduction
Campaign Health ScorerDevelopment
Detection Quality AgentDevelopment
Goal Progress MonitorDevelopment
Drift CompassDevelopment

Analysis and Intelligence ·13 sub-agents

Sub-agentStatus
Ask Maaten Query AgentProduction
Audience AnalystProduction
Creative AnalystProduction
Landing Page AgentProduction
Attribution AnalystDevelopment
Competitor AnalysisDevelopment
Incrementality AnalystDevelopment
Persona AnalystDevelopment
Marketing Effectiveness GuardianDevelopment
Video AnalyticsDevelopment
Web Analytics AgentDevelopment
Segmentation AnalystDevelopment
Benchmarks and TargetsDevelopment

Strategy, Forecasting, and Allocation ·9 sub-agents

Sub-agentStatus
BudgetProduction
Paid Media StrategistDevelopment
Incrementality Budget AllocatorDevelopment
Forecasting AgentDevelopment
Marketing Mix ModelerDevelopment
ML Predictive AgentDevelopment
CLV Prediction AgentDevelopment
Goal Feasibility AgentDevelopment
Campaign Expansion EvaluatorDevelopment

Creation and Execution ·7 sub-agents

Sub-agentStatus
Campaign BuilderProduction
Paid SearchProduction
Taxonomy GuardianDevelopment
Naming AnalyzerDevelopment
Cross-Platform TranslatorDevelopment
Geo Experimentation AgentDevelopment
Test DesignerPlanned

Coordination and Operations ·6 sub-agents

Sub-agentStatus
Account ManagerProduction
Head of DataDevelopment
Workflow DispatcherDevelopment
GTM CoordinatorDevelopment
GTM Strategic BrainDevelopment
Cross-Channel TeacherDevelopment

Control, security, and governance

Proprietary control layers.
Built into the architecture.

Security, governance, cost control, and access management are structural. Every layer is proprietary and enforced by the orchestration runtime.

Proprietary

Semantic Bouncers

Not rate-limiting. Not an off-the-shelf WAF. Maaten's proprietary firewall and AbuseGuard actively scan every input for jailbreaks, prompt injections, and toxic intent before the LLM ever processes it. Malicious prompts are blocked at the perimeter, not after inference.

Enforced

Human-in-the-Loop

For high-risk actions ·structural campaign changes, budget moves above threshold, mass communications ·the sub-agent queues an approval request. The action pauses until a human admin approves. No exceptions. No overrides.

Structural

Strict Tenant Isolation

Client data is physically segregated into specific directories. Multi-tenant infrastructure, isolated data. Organisations never share data, memory, or sub-agent state. Isolation is structural, not policy-based.

Complete

Full Audit Trail

Every sub-agent action logged: what happened, what data triggered it, which rule it followed, who approved it, what the outcome was. Traceable end to end. Built into the orchestration architecture from day one.

Configurable

Role-Based Access Control

Per organisation: who can hire sub-agents, who can approve actions, who can view reports, who can modify guardrails. Every permission boundary is enforced by the orchestration layer.

First-class

Token Cost Control

Per-agent, per-task token budgets with anomaly alerts. Intelligent routing to the most cost-effective model for each task. Caching layer for repeated queries. Anti-runaway-loop protection prevents recursive sub-agent calls from spiralling.

Earned, not licensed

Approved developer status.
Direct API access. Not through a 3rd party connector.

Maaten is an approved developer on Meta, Google, and TikTok. Raw data access and campaign push via secure, official API ·inside the channel relationship, not around it.

Live integrations

Meta Ads Google Ads TikTok Ads DV360 GA4 LinkedIn

Planned integrations

Shopify HubSpot Braze Mailchimp Amazon Ads Salesforce SAP Slack Looker Studio

Future channels

Programmatic (Roku, AWS) Connected TV Video / Audio Ads SEO data

Notification layer. Native Web Push API via VAPID keys for browser-level push notifications. Sub-agents can alert users directly without email or Slack dependency.

Named, owned, maintained

Proprietary data infrastructure.
Built and maintained in-house.

ComponentPurpose
Agentic EngineOrchestration runtime. Coordinates sub-agents, manages lifecycle, selects cognitive engine per task.
Data Uniformity EngineCross-platform metric standardisation. Consistent joining keys.
ELT PipelineCustom-built. Full control over extraction, loading, and transformation.
Gold TablesRefined, query-ready tables. 13+ business dimensions per campaign. Documented schemas.
Maaten IDUniversal joining key across platforms and the full data lifecycle.
Semantic MeshNatural language query layer. Powers "Ask Maaten." Plain English to structured queries.

See the Agent OS on your accounts.

Request a demo