WAF++ WAF++

Agentic – Scope

In Scope

The Agentic pillar covers all aspects of autonomous and semi-autonomous AI agent systems:

1. Agent Architectures

  • Single Agent: Individual autonomous agents with clear responsibilities

  • Swarm Architecture: Many homogeneous agents working in parallel

  • Mesh Architecture: Heterogeneous agents with specialized capabilities

  • Coordinator Pattern: Manager/worker patterns for orchestration

  • Hybrid Patterns: Combinations of above patterns

2. Agent Orchestration

  • Agent Coordination: How agents communicate and collaborate

  • Task Delegation: Distribution of work among agents

  • Result Aggregation: Combining outputs from multiple agents

  • Orchestration Tools: Step Functions, Lambda, Kubernetes controllers

3. Human-AI Interfaces

  • Human-in-the-Loop: Approval workflows for sensitive decisions

  • Human Supervision: Monitoring and intervention mechanisms

  • Feedback Loops: Human feedback for agent improvement

  • Alert Escalation: Handoff from agents to human operators

4. Agent Safety

  • Content Filtering: Blocking inappropriate or harmful content

  • Action Boundaries: Limiting what agents can do

  • Alignment: Ensuring agent goals align with organizational values

  • Fail-Safe Mechanisms: Automatic shutdown and recovery

5. Agent Observability

  • Metrics Collection: Token usage, latency, error rates

  • Tracing: End-to-end request tracking

  • Logging: Comprehensive audit trails

  • Dashboards: Visual monitoring and alerting

6. Agent Cost Management

  • Token Tracking: Input, output, and total token usage

  • Budget Alerts: Cost monitoring and warnings

  • Model Optimization: Cost-performance tradeoffs

  • Resource Efficiency: Efficient agent design

7. Agent Debt Management

  • Knowledge Cutoff Documentation: Version awareness

  • Tool Availability Gaps: Missing capabilities tracking

  • Accuracy Limitations: Known error patterns

  • Edge Cases: Unhandled scenarios documentation

  • Improvement Plans: Actionable debt resolution

Out of Scope

The following are explicitly out of scope for the Agentic pillar:

1. Prompt Engineering

  • Single prompt design and optimization

  • Prompt templates for chatbots

  • One-off prompt tuning

  • Related Pillar: Architecture/Operations

2. LLM Training

  • Model training and fine-tuning

  • Dataset curation for training

  • Training infrastructure management

  • Related Pillar: Architecture/Operations

3. Data Engineering

  • Training data preparation

  • Data pipeline infrastructure

  • Feature engineering for models

  • Related Pillar: Architecture/Operations

4. Infrastructure Provisioning

  • LLM model hosting infrastructure

  • GPU/TPU cluster management

  • Inference infrastructure scaling

  • Related Pillar: Reliability/Efficiency

5. General AI Safety

  • Model-level safety (bias, fairness)

  • General AI alignment research

  • Existential risk mitigation

  • Related Pillar: Security (if model-level)

Brownfield vs. Greenfield

Brownfield Agentic Implementation

  • Integration Pattern: Adding agents to existing systems

  • Migration Strategy: Incremental adoption with legacy compatibility

  • Orphaned Agents: Existing agents without governance

  • Integration Points: APIs, webhooks, message queues

Greenfield Agentic Implementation

  • Agent-First Design: Agents as core architectural components

  • Native Integration: No legacy constraints

  • Clean Boundaries: Clear separation of concerns

  • Optimized Flow: Designed for agentic patterns

Application Areas

The Agentic pillar applies to:

1. Customer-Facing Agents

  • Support Assistants: Helping customers with inquiries

  • Personal Assistants: Managing user schedules and tasks

  • Sales Assistants: Guiding purchasing decisions

  • Chatbots: Conversation management beyond simple FAQ

2. Internal Productivity Agents

  • Code Assistants: Writing and reviewing code

  • Documentation Assistants: Creating and updating docs

  • Meeting Assistants: Summarizing and action item extraction

  • Research Assistants: Information gathering and synthesis

3. Operational Agents

  • Monitoring Agents: Proactive issue detection

  • Response Agents: Automated incident response

  • Optimization Agents: Resource and cost optimization

  • Security Agents: Threat detection and response

4. Business Logic Agents

  • Decision Agents: Making business decisions within bounds

  • Workflow Agents: Orchestrating multi-step processes

  • Analysis Agents: Business intelligence and reporting

  • Compliance Agents: Regulatory compliance monitoring

Agent Deployment Stages

Stage What Happens Controls Applied

Discovery

Identify use cases for agents

WAF-AGN-100 (Agent Debt Register)

Design

Choose architecture pattern

WAF-AGN-060 (Multi-Agent Orchestration)

Development

Implement agent with guardrails

WAF-AGN-050 (Safety & Guardrails), WAF-AGN-020 (Tool Use Governance)

Deployment

Enable agent in production

WAF-AGN-010 (Identity & Authentication), WAF-AGN-080 (Human-in-the-Loop)

Operations

Monitor and optimize

WAF-AGN-070 (Agent Observability), WAF-AGN-090 (Cost & Performance)

Governance

Review and improve

WAF-AGN-040 (Memory & State), WAF-AGN-030 (Reasoning Transparency)

Scope Boundaries

Boundary Explanation

Tool Scope

Agents may only use tools explicitly defined in their configuration. Wildcard tool access is prohibited.

Data Scope

Agents may only access data they have been explicitly granted permission to access.

Action Scope

Agents may only perform actions within their defined permission boundaries.

Time Scope

Agent sessions and memory retention must be within defined time limits.

Interactions with Other Pillars

Pillar Interaction with Agentic

Operational Excellence

Agents require observability, runbooks, and postmortems like any production system.

Reliability

SLOs apply to agent responses. Agent failures are incidents requiring SRE response.

Security

Agent access must follow least privilege. Agent behavior must be monitored for compromise.

Cost

Agent operations represent cloud costs that must be tracked and optimized.

Architecture

Agent architecture decisions must follow architectural patterns and standards.

Governance

Agent debt register supports governance review. Controls map to compliance frameworks.

Sustainability

Agent compute impacts carbon footprint. Model selection affects environmental impact.

Efficiency

Agent efficiency (token usage, response time) is measured and optimized.