aws-strands

Strands Agents SDK

The open-source framework you write an agent's logic in - the "brain." Model-agnostic, AWS Bedrock by default. Runs anywhere (laptop, container, Lambda, EC2).

How this fits with the other AWS skill: Strands is the FRAMEWORK (what your agent does). To HOST and DEPLOY a Strands agent on AWS, use aws-harness (the AgentCore runtime). They compose: write with Strands, ship with AgentCore. You can also run Strands with no AWS deployment at all.

Install

pip install strands-agents strands-agents-tools

(A TypeScript SDK also exists - see the docs. Examples below are Python.)

Quick Start

from strands import Agent

agent = Agent()                       # defaults to Bedrock, Claude 4 Sonnet
print(agent("What is the capital of France?"))

agent(...) returns an AgentResult. str(result) gives the text; result.message is the structured dict (role + content).

Model configuration

Bedrock is the default provider and no model argument is needed - Strands picks a region-appropriate Claude 4 Sonnet. To override, pass a model id string or a BedrockModel provider:

from strands import Agent
from strands.models import BedrockModel

# Simple: a Bedrock model id (copy the exact id from the Bedrock model catalog)
agent = Agent(model="<your-bedrock-model-id>")

# Full control:
agent = Agent(model=BedrockModel(
    model_id="<your-bedrock-model-id>",
    temperature=0.3,
    region_name="us-west-2",
))

Strands is model-agnostic - other providers (Anthropic direct, OpenAI, etc.) are available via their own provider classes; see the model-providers docs.

Custom tools

The @tool decorator turns a function into something the model can call. The docstring is read by the model (first paragraph = description, Args: = parameter docs):

from strands import Agent, tool

@tool
def word_count(text: str) -> str:
    """Count the number of words in a piece of text.

    Args:
        text: The text to analyze.
    """
    return f"{len(text.split())} words"

agent = Agent(tools=[word_count])

Return recoverable strings on failure ("Error: ... ask the user to rephrase") instead of raising - the model reads the return value and can recover.

Prebuilt tools

from strands_tools import calculator   # from the strands-agents-tools package
agent = Agent(tools=[calculator])

Multi-agent patterns

Three verified patterns. Start with agents-as-tools (simplest delegation): wrap an agent in a @tool.

from strands import Agent, tool

researcher = Agent(system_prompt="You research topics thoroughly.")

@tool
def research(query: str) -> str:
    """Delegate a research question to the research specialist."""
    return str(researcher(query))     # str(AgentResult) = the text output

coordinator = Agent(tools=[research])
coordinator("Research the history of espresso and summarize it.")

For structured orchestration, use Swarm (agents hand off to each other dynamically) or Graph (a deterministic DAG where one node's output feeds the next):

from strands.multiagent import Swarm, GraphBuilder

# Swarm - dynamic handoffs
swarm = Swarm([researcher, writer, editor])
swarm("Draft and polish an article about espresso.")

# Graph - deterministic pipeline
builder = GraphBuilder()
builder.add_node(researcher, "research")
builder.add_node(writer, "write")
builder.add_edge("research", "write")     # research output -> writer input
graph = builder.build()
graph("Write an article about espresso.")

See the multi-agent docs for the full Graph/Swarm API.

Conversation management (context window)

Strands manages the conversation window for you (this is NOT long-term memory). The default is a sliding window:

from strands import Agent
from strands.agent.conversation_manager import SlidingWindowConversationManager

agent = Agent(conversation_manager=SlidingWindowConversationManager(window_size=20))

For durable, cross-session memory, use AgentCore Memory (see aws-harness) or the memory tools in strands-agents-tools.

Streaming

async for event in agent.stream_async("Explain quantum computing"):
    print(event)

Deploy on AWS

Strands runs anywhere. To put a Strands agent on AWS as a serverless endpoint with managed memory, identity, and observability, use the AgentCore harness: [aws-harness](../aws-harness/SKILL.md).

Resources

• Strands docs: https://strandsagents.com/
• Python quickstart: https://strandsagents.com/docs/user-guide/quickstart/python/
• Multi-agent patterns: https://strandsagents.com/docs/user-guide/concepts/multi-agent/multi-agent-patterns/
• Model providers: https://strandsagents.com/docs/user-guide/concepts/model-providers/
• GitHub: https://github.com/strands-agents/sdk-python