Kyndryl Agentic Framework Overview
Governance Foundation
The Kyndryl Agentic Framework (KAF) provides a comprehensive governance layer for enterprise AI applications. For Insurance Client's Use Case 1 (Unstructured Data Extraction & Quality Improvement), KAF ensures compliance, safety, observability, and operational excellence across the entire extraction pipeline.
KAF Purpose for UC1: Ensure the document extraction agent operates with full transparency, auditability, responsible AI practices, and seamless integration with enterprise governance requirements.
Why Enterprises Need More Than a Cluster
KAF vs Kubernetes
KAF is the Enterprise Control Plane for Agentic AI
addressing trust, safety, cost, and compliance gaps that Kubernetes does not natively solve.
KAF provides:
- Can we trust it
- control it
- explain it
- and run it safely at scale?
Every regulated or enterprise AI deployment needs KAF-level controls on top of Kubernetes.
| Dimension | Kubernetes (Baseline) | KAF – Agentic AI Platform |
|---|---|---|
| Primary Role | Container orchestration | ✓ Full agent runtime & governance layer |
| Agent Identity & Trust | ✗ Not native | ✓ Agent identity, certificates, mutual trust |
| Agent Lifecycle | ⚠ Pods/services only | ✓ Register, version, retire, hot-swap agents |
| Agent-to-Agent (A2A) | ✗ DIY messaging | ✓ A2A protocol, routing, guarantees |
| Tool Governance (MCP) | ✗ None | ✓ Tool catalog, permissions, audit |
| LLM Management | ✗ External scripts | ✓ Model routing, versioning, A/B testing |
| Token Economics | ✗ No cost visibility | ✓ Token budgets, attribution, alerts |
| Prompt Security | ✗ Not addressed | ✓ Injection & jailbreak protection |
| Output Guardrails | ✗ Not addressed | ✓ PII reduction, hallucination checks |
| Memory & State | ✗ Stateless by default | ✓ Short & long-term agent memory |
| RAG Infrastructure | ✗ DIY components | ✓ Governed embeddings & retrieval |
| Human-in-Loop | ✗ Custom build | ✓ Approval & escalation workflows |
| Explainability & Audit | ✗ Logs only | ✓ Decision traces & audit trails |
| Agent Observability | ⚠ Infra-level metrics | ✓ Per-agent traces, cost & DAGs |
| Multi-Tenancy | ⚠ Namespace-based | ✓ Tenant-isolated agents & policies |
| Graceful AI Degradation | ✗ Silent failures | ✓ Explicit AI failure handling |
| Responsible AI | ✗ Out of scope | ✓ Bias, fairness, policy enforcement |
| Testing & Simulation | ✗ Manual | ✓ Synthetic data & agent regression |
17 Core Capability Dimensions
Requirements Mapping
| Capability Dimension | UC1 Requirement | Implementation Status |
|---|---|---|
| 1. Agent Identity & Metadata | Document Extraction Agent card (/.well-known/agent.json) with capability declarations, versioning, owner info | Required |
| 2. Agent Lifecycle Management | Agent deployment, versioning, hot-swap capability for pipeline stages, graceful shutdown | Required |
| 3. Agent-to-Agent (A2A) Protocol | JSON-RPC 2.0 communication between upstream (Claims) and downstream (Underwriting, DW) systems | Required |
| 4. Tool Governance & Catalog | MCP-compliant tool discovery for Azure Document Intelligence, OpenAI, data validation tools | Required |
| 5. LLM Management & Routing | Multi-model support (GPT-4, GPT-3.5, Azure OpenAI), cost-aware routing, fallback strategies | Required |
| 6. Token Economics & Cost Control | Token usage tracking, cost per extraction, budget alerts, optimization recommendations | Required |
| 7. Prompt Management & Security | Versioned prompts, prompt injection protection, audit trail, role-based access to prompt changes | Required |
| 8. Output Validation & Guardrails | Schema validation, confidence scoring, automated escalation for low-confidence extractions, human-in-loop escalation | Required |
| 9. Memory & Context Management | Short-term context within a document, long-term patterns across policy corpus, conversation history | Required |
| 10. Retrieval-Augmented Generation (RAG) | Vector DB integration with Clause Indexer (Step 8) and Embedding (Step 9) for policy clause retrieval | Required |
| 11. Human-in-Loop Orchestration | Escalation workflows for ambiguous clauses, compliance reviewer sign-off, quality auditor validation | Required |
| 12. Explainability & Transparency | Decision audit trail showing extraction reasoning, clause citations, confidence scores | Required |
| 13. Observability & Monitoring | Azure Monitor integrations, Application Insights, real-time dashboards (95%, 60%, 80% KPIs), alert thresholds | Required |
| 14. Multi-Tenancy & Data Isolation | Insurance Client tenant isolation, policy data segregation, compliance with data residency requirements | Required |
| 15. Graceful Degradation | Fallback to manual extraction if agent fails, queue-based retry logic, exponential backoff | Required |
| 16. Responsible AI & Bias Detection | Policy fairness audits, extraction bias detection across document types, mitigation strategies | Required |
| 17. Testing & Simulation | Test coverage (unit/integration/e2e), synthetic document generation, chaos engineering for resilience | Required |
Detailed Capability Specifications
Implementation Guidance
1. Agent Identity & Metadata
Requirement: Document Extraction Agent must expose /.well-known/agent.json endpoint with complete capability declarations.
Implementation:
- Publish agent card with name, version, description, capabilities, owner, support contact
- Include tool discovery metadata (Azure Document Intelligence, OpenAI APIs)
- Versioning schema for API compatibility tracking
- Integration with Agent Registry for discoverability
2. Agent Lifecycle Management
Requirement: Support full lifecycle from deployment through retirement with zero-downtime updates.
Implementation:
- Containerized deployment (Docker) in Azure Kubernetes Service (AKS)
- Blue-green deployment for pipeline stage updates
- Graceful shutdown with in-flight request completion
- Version pinning for reproducibility across environments
3. Agent-to-Agent (A2A) Protocol
Requirement: Implement Google's A2A Open Protocol using JSON-RPC 2.0 for downstream integrations.
Implementation:
- HTTP/REST endpoints for Claims, Underwriting, Data Warehouse systems
- JSON-RPC 2.0 message format for method invocations
- Server-Sent Events (SSE) for streaming extraction results
- Request/response logging for audit trails
4. Tool Governance & Catalog
Requirement: Centralized tool discovery and governance for all extraction dependencies.
Implementation:
- Model Context Protocol (MCP) server for tool publication
- Tool versioning and deprecation policies
- Usage tracking and quota management
- Security scanning of tool implementations
5. LLM Management & Routing
Requirement: Intelligent routing across multiple LLMs with cost and latency optimization.
Implementation:
- Support for GPT-4, GPT-3.5, Azure OpenAI with dynamic model selection
- Cost-aware routing based on token estimates
- Latency-sensitive fallback chains
- Model performance benchmarking and A/B testing
6. Token Economics & Cost Control
Requirement: Full visibility into token usage and cost implications for every extraction.
Implementation:
- Per-document token tracking with cost attribution
- Budget alerts and spend forecasting
- Optimization recommendations for high-cost documents
- Financial reporting for chargeback models
7. Prompt Management & Security
Requirement: Secure, versioned prompt management with injection protection.
Implementation:
- Prompt version control with change history and rollback capability
- Injection attack detection and prevention
- Role-based access control for prompt modifications
- Compliance audit trail for regulatory requirements
8. Output Validation & Guardrails
Requirement: Automated validation with human escalation for low-confidence extractions.
Implementation:
- JSON schema validation against policy data model
- Confidence scoring (0-100%) for each extracted clause
- Automatic escalation for confidence < 85%
- Quality Checker (Step 6) integration with manual review queues
9. Memory & Context Management
Requirement: Short-term and long-term context to improve extraction accuracy.
Implementation:
- In-document context window for clause relationships
- Cross-document pattern learning from processed corpus
- Conversation history for refinement requests
- Redis cache for frequently accessed policy patterns
10. Retrieval-Augmented Generation (RAG)
Requirement: Vector DB integration for semantic clause retrieval and cross-policy similarity.
Implementation:
- Steps 8-10 pipeline: Clause Indexer → Embedding → Vector Store
- Azure Cognitive Search or Pinecone vector database
- Semantic similarity search for ambiguous clauses
- Policy precedent matching for decision support
11. Human-in-Loop Orchestration
Requirement: Workflow management for human review and approval gates.
Implementation:
- Escalation workflows triggered by confidence thresholds
- Compliance reviewer sign-off for sensitive clauses
- Quality auditor final validation before production release
- SLA-driven escalation (24-hour review cycle)
12. Explainability & Transparency
Requirement: Full audit trail showing extraction reasoning and confidence.
Implementation:
- Decision logs capturing LLM prompts and responses
- Source citation mapping extractions to original policy text
- Confidence scoring per clause with reasoning notes
- Audit interface for compliance teams to validate decisions
13. Observability & Monitoring
Requirement: Real-time dashboards tracking accuracy, speed, and error metrics.
Implementation:
- Azure Monitor integration with custom metrics
- Application Insights for end-to-end tracing
- KPI dashboards: 95% accuracy, 60% speed improvement, 80% error reduction
- Alerting thresholds for SLA violations
- Performance bottleneck identification and reporting
14. Multi-Tenancy & Data Isolation
Requirement: Insurance Client tenant isolation with strict data boundaries.
Implementation:
- Separate database schemas per tenant
- Azure Key Vault for tenant-specific credentials
- Query filtering and row-level security (RLS)
- Compliance with Australian data residency regulations
15. Graceful Degradation
Requirement: Fallback mechanisms for any pipeline stage failure.
Implementation:
- Automatic fallback to manual extraction if extraction agent fails
- Queue-based retry logic with exponential backoff
- Dead-letter queue for problematic documents
- Partial extraction capability (extract what's possible, flag what's not)
16. Responsible AI & Bias Detection
Requirement: Proactive detection and mitigation of extraction bias.
Implementation:
- Fairness audits across policy types and document formats
- Bias detection metrics for minority document categories
- Retraining triggers for detected bias patterns
- Compliance review for sensitive policy types (life, disability, etc.)
17. Testing & Simulation
Requirement: Comprehensive test coverage and resilience validation.
Implementation:
- Unit tests for each of the 13 pipeline steps
- Integration tests for A2A communication
- End-to-end testing with synthetic policy documents
- Chaos engineering for failure scenario validation
- Performance testing under peak 50+ document loads
KAF Integration with 13-Step Pipeline
Architecture Alignment
The Kyndryl Agentic Framework sits as a cross-cutting governance layer across all 13 extraction pipeline steps:
| Pipeline Step | KAF Dimension | Governance Focus |
|---|---|---|
| Step 1: Upload Handler | Agent Identity, Tool Governance | Request authentication, tool discovery, tenant isolation |
| Step 2: Document Identifier | Observability, Memory | Document classification logging, pattern learning |
| Step 3: OCR/Layout Extraction | Tool Governance, Output Validation | Azure Document Intelligence tool governance, confidence scoring |
| Step 4: Clause Parser | LLM Management, Token Economics | LLM routing, token tracking, cost attribution |
| Step 5: Policy Extractor | Prompt Management, Output Validation | Versioned extraction prompts, confidence scoring, escalation rules |
| Step 6: Quality Checker | Human-in-Loop, Explainability | Automated validation, escalation workflows, audit trails |
| Step 7: Data Normalizer | Output Validation, Responsible AI | Schema compliance, fairness checks, bias detection |
| Step 8: Clause Indexer | Tool Governance, Observability | Indexing tool governance, performance monitoring |
| Step 9: Embedding/Vectorization | LLM Management, Token Economics | Embedding model selection, cost tracking, usage quota |
| Step 10: Vector Store | Multi-Tenancy, Data Isolation | Tenant-specific vector indexes, access control |
| Step 11: Rules Transformer | Lifecycle Management, A2A Protocol | Rule versioning, downstream A2A communication prep |
| Step 12: Policy Store | Data Isolation, Graceful Degradation | Secure storage, failover mechanisms, backup/recovery |
| Step 13: Rule Evaluator/Decision Engine | A2A Protocol, Observability, Explainability | Downstream notifications, decision logging, audit trails |
Compliance & Governance
Risk Management
Regulatory Compliance
Applicable Standards:
- Australian Financial Services Licence (AFSL) requirements for Insurance Client
- Privacy Act 1988 (Cth) for personal information handling
- Consumer Law governance for policy terms clarity
- Internal AI Governance Framework alignment
Change Management
KAF Governance Controls:
- Prompt version control with peer review for changes
- Model swaps require technical and compliance sign-off
- Threshold changes (confidence limits, escalation rules) require business approval
- Monthly KAF compliance audits
Incident Response
KAF Capability Support:
- Explainability dimension provides rapid root-cause analysis
- Observability dimension enables quick detection of anomalies
- Graceful degradation ensures continuity during failures
- 2-hour incident response SLA with detailed postmortems
KAF Success Metrics
Performance Indicators
| KAF Dimension | Success Metric | Target Value |
|---|---|---|
| Observability | Extraction accuracy tracking | 95% target achieved |
| Observability | Processing speed improvement | 60% improvement vs. manual |
| Output Validation | Error reduction downstream | 80% reduction in claims/underwriting errors |
| Human-in-Loop | Escalation rate | < 5% of documents require manual review |
| Explainability | Audit trail completeness | 100% of decisions logged and traceable |
| LLM Management | Cost per extraction | < $0.50 per policy document |
| Testing | Test coverage | > 85% code coverage across 13 steps |
| Compliance | Audit findings | Zero critical findings in monthly audits |
Implementation Roadmap
Phased Approach
Phase 4 (Sprint): Foundation Setup
- ✓ Agent Identity card creation (/.well-known/agent.json)
- ✓ Basic observability dashboard (Accuracy, Speed, Errors)
- ✓ Output validation schema design
- ✓ Human-in-loop escalation workflow design
Phase 5 (POC Execution): Full Integration
- ☐ Deploy all 17 KAF capability dimensions
- ☐ A2A protocol implementation for downstream systems
- ☐ Real-time monitoring dashboards live in Azure Monitor
- ☐ Compliance audit framework operational
- ☐ 50+ document processing with full KAF governance
Phase 6 (Final Report): Optimization & Handover
- ☐ KAF compliance audit sign-off
- ☐ Governance playbooks for Insurance Client operations team
- ☐ Recommendations for ongoing KAF enhancements
- ☐ UC2 roadmap with KAF extension requirements
View Related Documents
← Return to Table of Contents for Sprint Plan, POC Specification, Architecture Diagram, and other project materials.