Strands Agents
by AWS
AWS open-source SDK (Apache 2.0) for building production-ready multi-agent systems. Model-driven approach where LLMs handle planning and orchestration. v1.0 released July 15, 2025. Supports Claude, GPT, Llama, and 10+ models. Features swarms, agent-to-agent handoffs, structured outputs, and observability via OpenTelemetry. 6,500+ GitHub stars; notable users: Smartsheet, Swisscom, Eightcap, Tavily.
Overview
Strands Agents reimagines multi-agent orchestration by letting the LLM decide how to orchestrate work — developer writes tools and business logic, LLM decides workflow. Designed from the ground up for production reliability, observability, and cost optimization.
Core philosophy: “The LLM handles planning and orchestration.”
Release Timeline
- May 2025: Initial preview launch
- July 15, 2025: Version 1.0 (production-ready)
- Growth: 6,500+ GitHub stars (Apache 2.0); 150K+ PyPI downloads at 1.0
Creators
- Organization: Amazon Web Services (AWS)
- Teams: Amazon Q Developer, AWS Glue, VPC Reachability Analyzer
- Leadership: Ryan Coleman (Product Manager), Belle Guttman (Engineering Lead)
Key Features
Multi-Agent Orchestration Patterns
Agents-as-Tools: Hierarchical delegation where agents can call other agents as tools
Handoffs: Human-in-the-loop workflow support with human transfer capabilities
Swarms: Self-organizing agent collaboration with emergent behavior (fully autonomous, dynamic)
Graphs: Deterministic, structured workflows with conditional logic and branching
Core Capabilities
- Model-driven planning: LLM-based task orchestration and agent selection
- Agent-to-Agent (A2A) protocol: Cross-platform agent communication
- Durable sessions: Automatic state persistence and recovery
- Async-first: Native concurrent execution
- Structured outputs: Automatic type validation and schema enforcement
- Resource management: Loop controls (max_iterations, max_execution_time, token_limit)
- Observability: Built-in OpenTelemetry tracing, logging, debugging
- Memory strategies: Multiple conversation memory management options
- Model switching: Mid-conversation model swaps for cost optimization
- Decorated functions: Python functions become tools with auto-generated descriptions
Model Support
Model-agnostic with native support for:
- Anthropic (Claude)
- OpenAI (GPT)
- Meta (Llama)
- Cohere
- Mistral
- Stability
- Writer
- Baseten
- Amazon Bedrock (native)
- Any OpenAI-compatible endpoint (LiteLLM)
- Local models
Architecture & Design
Model-Driven Approach: Developer defines structure (agents, tools, constraints); LLM decides how to orchestrate. Differs from graph-first frameworks where developer explicitly specifies flow.
Production Focus: Designed for reliability, observability, and resource constraints — not just prototype functionality.
Positioning vs Other Frameworks
vs LangGraph
| Aspect | Strands | LangGraph |
|---|---|---|
| Philosophy | Model-driven orchestration | Graph-first (developer decides) |
| API | Simpler, faster onboarding | More control, more boilerplate |
| Best for | Rapid development, dynamic reasoning | Complex workflows, fine-grained control |
| State | Implicit (LLM decides) | Explicit (reducer-based) |
vs CrewAI
- Strands: Production-grade reliability, observability, resource management
- CrewAI: Role-based agent patterns, simpler for specific use cases
- Strands edge: Better for high-volume deployments and cost-conscious ops
vs AutoGen
- AutoGen: In maintenance mode as of late 2025 (bug fixes only)
- Microsoft: Recommends Microsoft Agent Framework as successor
- Strands: Actively developed with regular feature releases
Use Cases
- Customer support with hierarchical routing and escalation
- Multi-team research and analysis (specialized agents per domain)
- Complex problem-solving requiring expertise delegation
- Tool orchestration with dynamic reasoning about tool selection
- Production deployments requiring conversation persistence
- Cost-optimized inference with mid-conversation model switching
- Human-in-the-loop workflows with approval gates
Enterprise Adoption
Notable Users:
- Smartsheet
- Swisscom
- Eightcap (financial services)
- Zafran
- Jit
- Tavily
- AutoScout24
Case Studies
Eightcap (Financial Services):
- Investigation time: 30 min → 45 seconds (94% improvement)
- Quality: 94% improvement in investigation quality
- Savings: $5M operational savings
- Deployment: 10 days (vs. traditional months-long development)
Amazon AMET Payments:
- Human-centric multi-agent approach with structured outputs
- Reduced hallucinations via schema enforcement
- Scaled across payment teams
AutoScout24:
- Standardized AI development framework for rapid, secure, scalable agent deployment
Licensing & Pricing
- License: Apache 2.0 (free, open-source)
- Cost model: Pay for underlying services (AWS, LLM API calls)
- No Strands-specific fees
When to Use Strands
✅ Choose Strands if:
- Need production-grade reliability and observability
- Want rapid agent development with model-driven orchestration
- Require multi-agent swarms with emergent behavior
- Cost optimization is critical (model switching, resource limits)
- High-volume deployments
- AWS ecosystem preference (Bedrock integration)
❌ Choose LangGraph if:
- Need fine-grained control over workflow orchestration
- Building complex multi-stage pipelines with explicit branching
- Financial/high-stakes approvals require auditability
- Prefer developer-explicit over LLM-driven decisions
Community
- GitHub: Open-source with active contribution
- Stars: 6,500+
- PyPI downloads: 150K+ (as of 1.0)
- Documentation: strandsagents.com