Model-agnostic support
Provides a unified interface to access multiple LLM providers including OpenAI, Anthropic, Google Vertex, Groq, and AWS Bedrock.
Type-safe validation
Uses Pydantic's validation system to ensure tool arguments and structured responses are validated, with errors returned to the LLM for retry.
Durable execution
Preserves agent progress across transient API failures, application errors, and restarts, supporting long-running and asynchronous workflows.
MCP and Agent-to-Agent integration
Supports the Model Context Protocol and Agent2Agent standards for external tool access and interoperability between agents.
Human-in-the-loop tool approval
Allows certain tool calls to require approval before execution based on tool arguments, conversation history, or user preferences.
Streamed outputs
Enables continuous streaming of structured output with immediate validation for real-time data access.