Features
Any model from any provider
Lifecycle hooks (BeforeToolCallEvent, etc.)
Custom tool definitions with Zod (TS) or decorators (Python)
Multi-agent orchestration
Strands Shell sandboxed execution
Strands Evals integration
Context management reduces costs by 50%
Chaos testing for resilience
MLLM-as-a-Judge evaluators (image-to-text)
Structured outputs with type safety
Steering hooks for deterministic guardrails
MCP server for coding agents (Claude, Cursor, etc.)
Observability without config
Multi-agent orchestration with handoffs
Sandbox Agents for containerized long-running tasks
Agent-as-tool delegation (v0.15.0+)
Realtime Agents with gpt-realtime-2 voice support (v0.17.6+)
Human-in-the-loop mechanisms
Automatic session history management
Built-in tracing for debugging
Provider-agnostic LLM support (100+ models via LiteLLM)
Redis session support (optional, v0.17.6+)
Instructions, tools, and guardrails configuration
Jupyter notebook compatibility
Supports OpenAI Responses and Chat Completions APIs
