Maturity Model: Agentic Maturity

The maturity model of the Agentic pillar assesses the organizational and technical maturity of autonomous AI agents across five levels — from ungoverned experiments to self-monitoring, optimized agent systems.

The Five Maturity Levels

Level	Name	Description
1	Ad-hoc / Experimental	Agents are built in isolation and deployed to production without governance. No stable identities, no central logs, no guardrails. Failures and costs are not traceable.
2	Documented	Initial agent architectures and tool lists are documented. Agents have names and simple IDs; logs are reviewed manually. Basic guardrails and oversight processes are defined, but not yet consistently enforced.
3	Governed	All ten Agentic controls are implemented and measurable. Stable agent IDs, tool governance, reasoning traces, safety guardrails, human-in-the-loop, and observability are standard. Training and quarterly reviews are established.
4	Optimized	Cost and performance monitoring are active; automated model selection, caching, and scaling. Multi-agent orchestration with failover. The Agent Debt Register is linked to the development backlog. Regular alignment tests are run.
5	Autonomous / Self-Improving	AI-driven optimization of the controls themselves. Predictive guardrails, adaptive human-in-the-loop, and self-healing orchestration. Measurable reduction of agent debt and incident likelihood over time.

Level

Name

Description

Ad-hoc / Experimental

Agents are built in isolation and deployed to production without governance. No stable identities, no central logs, no guardrails. Failures and costs are not traceable.

Documented

Initial agent architectures and tool lists are documented. Agents have names and simple IDs; logs are reviewed manually. Basic guardrails and oversight processes are defined, but not yet consistently enforced.

Governed

All ten Agentic controls are implemented and measurable. Stable agent IDs, tool governance, reasoning traces, safety guardrails, human-in-the-loop, and observability are standard. Training and quarterly reviews are established.

Optimized

Cost and performance monitoring are active; automated model selection, caching, and scaling. Multi-agent orchestration with failover. The Agent Debt Register is linked to the development backlog. Regular alignment tests are run.

Autonomous / Self-Improving

AI-driven optimization of the controls themselves. Predictive guardrails, adaptive human-in-the-loop, and self-healing orchestration. Measurable reduction of agent debt and incident likelihood over time.

Control Maturity Matrix

This matrix maps each WAF-AGN control to the target maturity per level:

Control	Level 1	Level 2	Level 3	Level 4	Level 5
WAF-AGN-010 Agent Identity & Authentication	No identity	Names, manual logs	Stable IDs, IAM/service account, audit trail	SIEM integration, anomaly detection	Behavioral baselining, identity anomalies
WAF-AGN-020 Tool Use Governance	All tools allowed	Some tools logged	Whitelist, rate limiting, tool schemas	Automated policy monitoring	Predictive abuse detection
WAF-AGN-030 Reasoning Transparency	Black box	Input/output logged	Chain-of-thought, tool calls, decision logic	Post-mortem and pattern analysis	Predictive reasoning optimization
WAF-AGN-040 Memory & State Management	No memory	Simple storage	Versioned memory, state backups	Automated management, real-time backups	Predictive memory management
WAF-AGN-050 Safety & Guardrails	No guardrails	Some content filters	Full filters, action boundaries, HIL	Automated compliance enforcement	Predictive guardrails
WAF-AGN-060 Multi-Agent Orchestration	No orchestration	Manual coordination	Event-based coordination, basic failover	Dynamic task assignment, auto-failover	Self-healing orchestration
WAF-AGN-070 Agent Observability	No observability	Basic logging	Token, latency, error metrics, dashboards	Proactive monitoring, auto-scaling	Predictive observability
WAF-AGN-080 Human-in-the-Loop	No HIL	Manual approvals	Workflows for sensitive decisions, audit trail	Automated decision classification	Adaptive HIL
WAF-AGN-090 Cost & Performance Monitoring	No monitoring	Manual cost tracking	Automated token tracking, budget alerts	Intelligent model selection, caching	AI-based cost-performance optimization
WAF-AGN-100 Agent Debt Register	No register	Ad-hoc documentation	Formal register, quarterly reviews	Issue-tracking integration, dashboards	Predictive debt management

Control

Level 1

Level 2

Level 3

Level 4

Level 5

WAF-AGN-010 Agent Identity & Authentication

No identity

Names, manual logs

Stable IDs, IAM/service account, audit trail

SIEM integration, anomaly detection

Behavioral baselining, identity anomalies

WAF-AGN-020 Tool Use Governance

All tools allowed

Some tools logged

Whitelist, rate limiting, tool schemas

Automated policy monitoring

Predictive abuse detection

WAF-AGN-030 Reasoning Transparency

Black box

Input/output logged

Chain-of-thought, tool calls, decision logic

Post-mortem and pattern analysis

Predictive reasoning optimization

WAF-AGN-040 Memory & State Management

No memory

Simple storage

Versioned memory, state backups

Automated management, real-time backups

Predictive memory management

WAF-AGN-050 Safety & Guardrails

No guardrails

Some content filters

Full filters, action boundaries, HIL

Automated compliance enforcement

Predictive guardrails

WAF-AGN-060 Multi-Agent Orchestration

No orchestration

Manual coordination

Event-based coordination, basic failover

Dynamic task assignment, auto-failover

Self-healing orchestration

WAF-AGN-070 Agent Observability

No observability

Basic logging

Token, latency, error metrics, dashboards

Proactive monitoring, auto-scaling

Predictive observability

WAF-AGN-080 Human-in-the-Loop

No HIL

Manual approvals

Workflows for sensitive decisions, audit trail

Automated decision classification

Adaptive HIL

WAF-AGN-090 Cost & Performance Monitoring

No monitoring

Manual cost tracking

Automated token tracking, budget alerts

Intelligent model selection, caching

AI-based cost-performance optimization

WAF-AGN-100 Agent Debt Register

No register

Ad-hoc documentation

Formal register, quarterly reviews

Issue-tracking integration, dashboards

Predictive debt management

Self-Assessment Checklist: Level 2 (Documented)

An organization is at Level 2 when all of the following are true:

All production agents are captured in a central inventory with name, owner, and purpose
At least one agent has a stable ID and a documented authentication concept
Agent logs (input/output) are collected for all production agents
An initial tool whitelist or approval list is documented
Basic guardrails (e.g., content filters) are enabled for at least one production agent
A documented process exists for reviewing sensitive agent decisions

Self-Assessment Checklist: Level 3 (Governed)

An organization is at Level 3 when all of the following are true:

All ten WAF-AGN controls have been assessed and implemented at least at Level 3
Every agent has a unique, stable ID and authenticates to tools and other agents
All tool usage is logged; rate limiting and tool schemas are configured
Chain-of-thought and tool calls are logged for all critical decisions
Memory and state are versioned or persisted; backups are regularly tested
Content filters, action boundaries, and human-in-the-loop are active for sensitive actions
Multi-agent systems use event-based communication with retry and DLQ logic
Token, latency, and error metrics are visible in dashboards and have alerting
Budget alerts for agent costs are configured at 50%, 75%, 90%, and 100%
An Agent Debt Register exists and was reviewed in the last quarter

Self-Assessment Checklist: Level 4 (Optimized)

An organization is at Level 4 when all of the following additionally apply:

Agent logs flow automatically into SIEM or central log analysis
Tool usage is automatically checked for policy violations and escalated
Reasoning traces enable post-mortem analysis within 24 hours
Memory-state backups and point-in-time recovery are active for production stores
Guardrails are regularly tested (at least quarterly) against organizational values
Multi-agent orchestration uses dynamic task assignment and automatic failover
Latency, error, and budget anomalies trigger automated alerts and runbooks
Model selection and caching are optimized per task (cost-performance tradeoff)
The Agent Debt Register is linked to issue tracking; progress is reported quarterly

Recommended Entry Path

Phase	Timeframe	Actions	Target Level
Phase 0: Inventory	Week 1–2	Capture all production agents, inventory IDs and tool access, create risk rating	Level 1 → 2
Phase 1: Governance Foundation	Month 1–3	Implement WAF-AGN-010/020/050/080: IDs, auth, tool whitelist, guardrails, HIL	Level 2 → 3
Phase 2: Transparency & State	Month 2–4	Reasoning logging (030), versioned memory (040), observability (070), budget alerts (090)	Level 3
Phase 3: Scaling	Quarter 2–4	Multi-agent orchestration (060), auto-scaling, caching, model-selection optimization	Level 3 → 4
Phase 4: Continuous Improvement	Ongoing	Agent Debt Register (100), alignment tests, predictive guardrails, self-optimization	Level 4 → 5

Phase

Timeframe

Actions

Target Level

Phase 0: Inventory

Week 1–2

Capture all production agents, inventory IDs and tool access, create risk rating

Level 1 → 2

Phase 1: Governance Foundation

Month 1–3

Implement WAF-AGN-010/020/050/080: IDs, auth, tool whitelist, guardrails, HIL

Level 2 → 3

Phase 2: Transparency & State

Month 2–4

Reasoning logging (030), versioned memory (040), observability (070), budget alerts (090)

Level 3

Phase 3: Scaling

Quarter 2–4

Multi-agent orchestration (060), auto-scaling, caching, model-selection optimization

Level 3 → 4

Phase 4: Continuous Improvement

Ongoing

Agent Debt Register (100), alignment tests, predictive guardrails, self-optimization

Level 4 → 5

Most organizations start at Level 1 or 2. Level 3 is the operational minimum for production agents. Level 4 is the target for scaled, business-critical agentic systems. Level 5 remains a medium-term goal (3–5 years) for most organizations.