kaizen

Define a Signature subclass with InputField and OutputField, create a dataclass config with llm_provider and model, then extend BaseAgent passing config and signature to super().__init__(). Call agent.run() with your inputs to execute.

Create a SharedMemoryPool, instantiate agents with agent_id values sharing that pool, and coordinate them with patterns like SupervisorWorkerPattern or Pipeline.router. The A2A protocol provides automatic capability cards and semantic agent matching.

The unified Agent class uses agent_type to select execution patterns: 'simple' for single inference, 'cot' for chain-of-thought reasoning, 'react' for reason-plus-act cycles with tools, and 'rag' for retrieval-augmented generation. Each preset configures cycles, tools, and memory automatically.

VisionAgent requires file paths instead of base64 strings when using Ollama, and accepts a 'question' parameter rather than 'prompt'. Results are returned under the 'answer' key, not 'response'. Read the multimodal pitfalls guide before using vision or audio.

Yes, the Enterprise Agent Trust Protocol (EATP) provides cryptographic trust chains, TrustedAgent classes, HMAC-authenticated secure messaging, credential rotation, and audit logging. It also supports RBAC and compliance controls for SOC2, GDPR, HIPAA, and PCI-DSS.

No, Kaizen recommends a 3-tier testing strategy using real infrastructure in Tiers 2 and 3 (Ollama locally, then OpenAI). Mocking LLM calls in integration tests is explicitly discouraged because it hides real provider behavior.