Kaizen Framework Skills

27 skills for Kaizen AI agent framework covering core patterns, multi-agent coordination, multi-modal processing, autonomous tool calling, observability, memory systems, checkpoint/resume, and journey orchestration.

Skill Categories

Core Patterns (6 Skills - CRITICAL/HIGH)

Essential patterns for building Kaizen agents:

kaizen-baseagent-quick.md - BaseAgent implementation, signature, config (3-step pattern)
kaizen-signatures.md - InputField, OutputField, type-safe I/O, validation
kaizen-config-patterns.md - Domain config vs BaseAgentConfig, auto-extraction
kaizen-ux-helpers.md - extract_list/dict/float/str(), write_to_memory()
kaizen-agent-execution.md - agent.run(), result handling, async execution
kaizen-quickstart-template.md - Complete agent template (copy-paste ready)

Quick Start: Begin with kaizen-baseagent-quick.md → kaizen-signatures.md → kaizen-quickstart-template.md

Multi-Agent (5 Skills - HIGH)

Multi-agent coordination and Google A2A protocol:

kaizen-multi-agent-setup.md - SharedMemoryPool, agent coordination infrastructure
kaizen-shared-memory.md - write_to_memory(), read_relevant(), patterns
kaizen-a2a-protocol.md - Automatic capability cards, semantic matching (100% Google A2A)
kaizen-supervisor-worker.md - Supervisor-worker pattern, task delegation
kaizen-agent-patterns.md - Consensus, debate, specialists, producer-consumer

Multi-Modal (4 Skills - HIGH)

Vision, audio, and multi-modal processing:

kaizen-vision-processing.md - VisionAgent, OllamaVisionProvider, image analysis
kaizen-audio-processing.md - Whisper, audio transcription
kaizen-multimodal-orchestration.md - MultiModalAgent, vision+audio+text
kaizen-multimodal-pitfalls.md - CRITICAL: Common mistakes (kaizen-specialist:301-373)

IMPORTANT: Read kaizen-multimodal-pitfalls.md FIRST to avoid common API mistakes

Journey Orchestration (1 Skill - HIGH) - NEW in v0.9.0

User journey management with declarative pathways and intent-driven transitions:

kaizen-journey-orchestration.md - Journey, Pathway, Transitions, Context Accumulation, Nexus deployment

Key Features: Multi-pathway flows, intent detection (LLM-powered), ReturnToPrevious behavior, context accumulation with merge strategies, Nexus deployment

Advanced Patterns (11 Skills - MEDIUM/HIGH)

Production patterns, enterprise features, tool calling, observability, memory systems, and specialized techniques:

kaizen-control-protocol.md - Bidirectional agent ↔ client communication
kaizen-tool-calling.md - Autonomous tool execution with approval workflows
kaizen-observability.md - Complete observability stack (tracing, metrics, logging, audit)
kaizen-memory-system.md - Persistent memory, learning, and preference adaptation
kaizen-checkpoint-resume.md - Automatic checkpointing and resume for long-running agents
kaizen-chain-of-thought.md - CoT pattern, step-by-step reasoning
kaizen-rag-agent.md - RAG implementation with Kaizen
kaizen-react-pattern.md - ReAct (reasoning + acting)
kaizen-cost-tracking.md - Token usage, budget management
kaizen-streaming.md - Streaming responses, real-time output
kaizen-testing-patterns.md - 3-tier testing, fixtures, standardized tests

Testing: All patterns use 3-tier strategy (Unit → Ollama → OpenAI), Real infrastructure recommended in Tiers 2-3

Learning Paths

Path 1: Basic Agent (15 minutes)

kaizen-baseagent-quick.md - Core pattern
kaizen-signatures.md - I/O definitions
kaizen-quickstart-template.md - Copy template and run

Output: Working Q&A agent

Path 2: Production Agent (30 minutes)

kaizen-baseagent-quick.md - Core pattern
kaizen-config-patterns.md - Production config
kaizen-ux-helpers.md - Defensive parsing
kaizen-agent-execution.md - Error handling
kaizen-testing-patterns.md - 3-tier testing

Output: Production-ready agent with tests

Path 3: Multi-Agent System (45 minutes)

kaizen-baseagent-quick.md - Core pattern
kaizen-multi-agent-setup.md - Infrastructure
kaizen-shared-memory.md - Coordination
kaizen-a2a-protocol.md - Semantic matching
kaizen-supervisor-worker.md - Task delegation

Output: Multi-agent system with semantic routing

Path 4: Multi-Modal Agent (30 minutes)

kaizen-baseagent-quick.md - Core pattern
kaizen-multimodal-pitfalls.md - READ FIRST!
kaizen-vision-processing.md - Image analysis
kaizen-audio-processing.md - Audio transcription
kaizen-multimodal-orchestration.md - Unified processing

Output: Vision + audio agent

Path 5: User Journey Orchestration (45 minutes) - NEW

kaizen-baseagent-quick.md - Core pattern
kaizen-signatures.md - Signatures with __intent__, __guidelines__
kaizen-journey-orchestration.md - Journey, Pathway, Transitions
examples/journey/healthcare_referral/ - Reference implementation

Output: Multi-pathway user journey with intent-driven transitions

Critical References

Quick References

Specialist Agent: .claude/agents/frameworks/kaizen-specialist.md (comprehensive reference table)
Examples: the Kaizen examples (35+ working examples)

Key Content Sources

Multi-Modal Pitfalls: kaizen-specialist.md lines 301-373 (CRITICAL)
A2A Protocol: kaizen-specialist.md lines 115-165
UX Improvements: kaizen-specialist.md lines 249-298
Quickstart Template: kaizen-specialist.md lines 489-520
Test Fixtures: tests/conftest.py

Critical Patterns

BaseAgent Pattern (Most Common)

from kaizen.core.base_agent import BaseAgent
from kaizen.signatures import Signature, InputField, OutputField
from dataclasses import dataclass

@dataclass
class MyConfig:
    llm_provider: str = os.environ.get("LLM_PROVIDER", "openai")
    model: str = os.environ.get("LLM_MODEL", "")
    temperature: float = 0.7

class MySignature(Signature):
    question: str = InputField(description="User question")
    answer: str = OutputField(description="Answer")

class MyAgent(BaseAgent):
    def __init__(self, config: MyConfig):
        super().__init__(config=config, signature=MySignature())

    def ask(self, question: str) -> dict:
        return self.run(question=question)

Multi-Agent Pattern

from kaizen.memory.shared_memory import SharedMemoryPool

shared_pool = SharedMemoryPool()

agent1 = ResearcherAgent(config, shared_pool, agent_id="researcher")
agent2 = AnalystAgent(config, shared_pool, agent_id="analyst")

findings = agent1.research("AI trends")
analysis = agent2.analyze(findings)

Vision Pattern (Watch for Pitfalls!)

from kaizen_agents.agents import VisionAgent, VisionAgentConfig

config = VisionAgentConfig(llm_provider="ollama", model="bakllava")
agent = VisionAgent(config=config)

result = agent.analyze(
    image="/path/to/image.png",  # File path, NOT base64
    question="What is this?"     # 'question', NOT 'prompt'
)
print(result['answer'])          # Key is 'answer', NOT 'response'

Framework Status (v0.9.0)

Implementation: Production-ready Performance: -0.06% overhead (essentially zero), validated with real LLM workloads Observability: Complete stack (tracing, metrics, logging, audit) with Grafana/Prometheus/Jaeger/ELK Multi-Modal: Vision (Ollama + OpenAI) + Audio (Whisper) fully operational Multi-Agent: SupervisorWorkerPattern production-ready A2A Protocol: 100% Google A2A compliant with automatic capability cards Journey Orchestration: Layer 5 with declarative pathways, intent detection, context accumulation (351 tests)

CRITICAL RULES

ALWAYS:

✅ Use domain configs (e.g., QAConfig), let BaseAgent auto-convert
✅ Call self.run(), not strategy.execute()
✅ Load .env with load_dotenv() before creating agents
✅ Use extract_*() for result parsing
✅ Read kaizen-multimodal-pitfalls.md before using vision/audio

NEVER:

❌ Create BaseAgentConfig manually (use auto-conversion)
❌ Use 'prompt' parameter with VisionAgent (use 'question')
❌ Pass base64 strings to Ollama (use file paths)
❌ Access 'response' key from VisionAgent (use 'answer')
❌ Skip real infrastructure testing (Real infrastructure recommended in Tiers 2-3)

Quick References by Task

Task	Skills
Create basic agent	baseagent-quick, signatures, quickstart-template
Production agent	config-patterns, ux-helpers, agent-execution, testing-patterns
Production monitoring	observability (v0.5.0) - tracing, metrics, logging, audit
Interactive agents	control-protocol (v0.2.0), baseagent-quick
Tool calling	tool-calling (v0.2.0), control-protocol
Multi-agent system	multi-agent-setup, shared-memory, a2a-protocol, supervisor-worker
User journeys	journey-orchestration (v0.9.0) - pathways, intent detection, context
Vision processing	multimodal-pitfalls (READ FIRST), vision-processing
Audio processing	audio-processing, multimodal-orchestration
Chain of thought	chain-of-thought
RAG implementation	rag-agent
Cost management	cost-tracking
Streaming	streaming

Next Steps: Start with kaizen-baseagent-quick.md for core pattern, then choose a learning path based on your needs.