Arsanjani GenAI Maturity Model

Overview

The GenAI Maturity Model from Agentic Architectural Patterns for Building Multi-Agent Systems (Arsanjani & Bustos, Packt, 2026) defines seven levels (0–6) as a strategic roadmap for enterprise AI adoption — from raw data preparation to sophisticated multi-agent collaboration. It is notable for explicitly mapping maturity levels to architectural patterns: an organization's maturity is a direct result of the specific patterns it implements, not just its stated intent.

Security, Privacy, and Compliance are treated as cross-cutting concerns at every level.

Maturity Levels

Level	Title	Key Activities	Example
0	Prepare data (data foundation)	Acquire, generate (including synthetic), clean, curate, and govern data; address quality, relevance, licensing, and accessibility	Building a governed data lake; establishing data lineage
1	Select model and prompt/serve	Select pretrained foundation models; design prompts (prompt engineering); deploy and serve via APIs for content generation, Q&A, and basic function calling	A chatbot that answers HR questions using model knowledge
2	Contextual enhancement (RAG)	Use RAG to dynamically fetch external knowledge (documents, databases) and augment prompts; improves accuracy and reduces hallucination	An internal knowledge bot that retrieves live policy documents before answering
3	Tuning for specificity	Fine-tune models using PEFT/LoRA adaptor tuning or full fine-tuning with domain-specific data; specializes terminology, style, and behavior for agent roles	Tuning a model on enterprise sales data to understand sales-specific jargon
4	Grounding and evaluation	Implement output grounding (citations, source linking) and robust evaluation frameworks covering accuracy, bias, fairness, and safety	A financial analysis agent that provides summaries with explicit references to source reports
5	Single-agent systems	Architect systems around one autonomous AI agent performing multi-step reasoning, planning, and tool use (via function calling or MCP); requires LLMOps/AgentOps for lifecycle management	A travel planning agent that autonomously books flights and hotels via real APIs
6	Multi-agent systems	Deploy multiple specialized agents that collaborate, coordinate (via A2A protocol), and negotiate to solve problems exceeding a single agent's capacity	A supply chain system where inventory, logistics, and forecasting agents collaborate to respond to disruptions

Agentic AI Maturity Spectrum (Expanded Levels 5–6)

Within the agentic portion (Levels 5–6), the book defines six finer-grained sub-levels that map directly to the coordination pattern catalog:

Sub-level	Description	Key Patterns
1 – Basic agentic	Single agents, fixed workflows, predefined tool calls	Single-Agent Baseline, Static Function Calling, Watchdog Timeout, Agent Calls Human
2 – Dynamic single-agent	Single agent dynamically selects from pre-approved tools	Agent Router, Dynamic Tool Selection, Simple RAG, Simple Retry
3 – Introspective (ReAct/Reflexion)	Self-reflection and self-correction feedback loops	ReAct, Reflexion, Instruction Fidelity Auditing, Adaptive Retry with Prompt Mutation
4 – Multi-agent systems	Multiple specialized agents; structured top-down coordination	Supervisor Architecture, Multi-Agent Planning, Shared Epistemic Memory, Event-Driven Reactivity, Tool/Agent Registry
5 – Advanced with meta-agents	A meta-agent oversees and dynamically reassigns work	Meta-agents, Blackboard Topology, Resource Allocation, Contract-Net Marketplace, Supervision Trees
6 – Self-correcting agents	Multi-turn feedback loops; agents critique and refine each other iteratively	Consensus, Agent Negotiation, Conflict Resolution, FCoT Embedding, Coevolved Agent Training, Trust Decay

The New Agentic Stack

Advancing toward Levels 5–6 requires mastering three complementary interoperability layers:

Layer	Technology	Purpose
1 – Function calling	Native LLM capability	Agent's LLM invokes local tools within a single application runtime
2 – Model Context Protocol (MCP)	Anthropic's open standard	Standardizes how agents discover and invoke external tools as independent interoperable services (agent → tools)
3 – Agent-to-Agent (A2A)	Google's open protocol	Universal standard for structured task delegation and collaboration between independent agents on any framework (agent → agent)

Mental model: MCP = agent connects to tools; A2A = agents connect to each other. In a complete multi-agent workflow, an orchestrator uses A2A to delegate tasks → worker agents use MCP to access the tools they need → results flow back through A2A.

Agent Anatomy

The book defines seven internal components that form the continuous operational loop of any AI agent:

Component	Function	Implementation
Goals	Objectives the agent seeks to achieve	Configuration parameters or dynamic state
Sense (Perception)	Gathers data from environment (APIs, databases, sensors)	API listeners, data stream processors, MCP clients
Reason (Cognition)	Analyzes sensed information using LLM	The agent-ready LLM; interprets inputs against goals
Plan	Devises a sequence of actions	LLM-generated task sequence; can be static or dynamic
Act (Action)	Executes plan via tools	External API calls, code execution, response generation
Memory	Stores knowledge, state, and past experience	Short-term (in-context); long-term (vector databases, persistent stores)
Coordinate	Interacts with other agents (multi-agent systems only)	A2A protocol; tracks task lifecycle states (submitted, working, completed)

The agentic loop: Sense → Reason → Plan → Act → (feedback) → Sense. Each iteration allows the agent to adapt based on outcomes.

Practical Rollout Roadmap

Chapter 12 of the book maps patterns to three implementation stages:

Stage	Core Principle	Patterns to Implement
1 – Foundational system	Build one thing well; test every integration	Single-Agent Baseline, basic observability, static tool calling
2 – Production-ready service	Reliability and scalability	Watchdog Timeout, Adaptive Retry, Canary Testing, checkpointing, evaluation pipelines
3 – Self-improving ecosystem	Continuous learning and multi-agent coordination	Coevolved Agent Training, Trust Decay, Swarm Architecture, Consensus

Key Production Challenges

The book identifies four categories of challenges in transitioning from PoC to production:

Category	Key Challenges
Strategic and organizational	Demonstrating ROI, achieving stakeholder alignment, operational integration, change management
Data-related	Data governance, data quality, data silos, privacy compliance (GDPR, HIPAA)
Model and technical	Model robustness, adversarial attacks, hallucination minimization, legacy system integration, LLMOps/AgentOps
Ethical and responsible AI	Bias mitigation, transparency, explainability, governance frameworks

References

Arsanjani, A., & Bustos, J.P. (2026). Agentic Architectural Patterns for Building Multi-Agent Systems. Packt Publishing. ISBN 978-1-80602-957-0. — Comprehensive pattern catalog and GenAI maturity model; code at https://github.com/PacktPublishing/Agentic-Architectural-Patterns-for-Building-Multi-Agent-Systems