Strands Agents SDK

Warnings

• breaking A critical supply chain attack impacted 'litellm' (a dependency) versions 1.82.7 and 1.82.8, which contained malicious code. Strands Agents v1.33.0 (and later) hard-pinned `litellm<=1.82.6` to mitigate this. Users on older Strands versions or with unpinned `litellm` dependencies should immediately upgrade `strands-agents` and audit their environments.
  Fix: Upgrade `strands-agents` to `1.33.0` or higher. If affected, follow `litellm`'s security advice for credential rotation and system audit.
• gotcha Strands Agents defaults to using Amazon Bedrock for models. This requires AWS credentials (e.g., `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, `AWS_REGION`) to be configured in your environment or AWS CLI, and model access to be enabled in your AWS account. Without proper configuration, agents may fail to initialize or execute.
  Fix: Ensure AWS credentials are correctly configured and the necessary Bedrock models are enabled in your AWS account. For non-Bedrock models, explicitly configure the desired `model` provider during `Agent` initialization.
• gotcha The `pyaudio` dependency is optional and only included with the `bidi-io` extra. If you intend to use bidirectional streaming features requiring audio input/output, you must install `strands-agents` with this extra (e.g., `pip install strands-agents[bidi,bidi-io]`).
  Fix: Install the `bidi-io` extra: `pip install strands-agents[bidi,bidi-io]`.
• gotcha When using `BedrockModel`, specifying an unsupported `service_tier` (e.g., 'flex', 'priority') for a given model or region will result in a `ValidationException`. Not all models or AWS regions support all service tiers.
  Fix: Consult AWS Bedrock documentation for supported `service_tier` values per model and region, or omit the `service_tier` parameter to use the default behavior.
• gotcha When granting agents access to tools that perform sensitive actions (e.g., SQL execution, file system operations, API calls), implement robust input validation, output sanitization, and least-privilege IAM policies. Relying solely on system prompts for security is insufficient against prompt injection or unforeseen agent behavior in production environments, potentially leading to data leakage or destructive actions.
  Fix: Implement custom middleware or 'steering hooks' to inspect and validate tool inputs/outputs, use fine-grained IAM roles for tool execution, and thoroughly test agents in controlled environments.

Install

• pip install strands-agents strands-agents-tools Core SDK and recommended tools
• pip install strands-agents[bidi,bidi-io] For bidirectional streaming with audio I/O (includes PyAudio)

Imports

• Agent

  from strands import Agent

• tool

  from strands import tool

  Used as a decorator to define custom tools.
• calculator

  from strands_tools import calculator

  Example tool from the 'strands-agents-tools' package.
• BedrockModel

  from strands.models.bedrock import BedrockModel

  For explicit configuration of Amazon Bedrock models.

Quickstart

This quickstart demonstrates creating a basic Strands agent with a pre-built tool (`calculator`) from `strands-agents-tools` and querying it. Ensure AWS credentials are configured as Strands uses Amazon Bedrock as the default model provider.

import os
from strands import Agent
from strands_tools import calculator

# Configure AWS credentials for Bedrock (if not already set via environment variables)
# For example, by setting AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY
# Or ensure your AWS CLI is configured with a default region and profile

agent = Agent(tools=[calculator])

async def run_agent():
    print("Agent initialized. Asking a question...")
    response = await agent("What is the square root of 1764?")
    print(f"Agent response: {response.content}")

if __name__ == "__main__":
    import asyncio
    asyncio.run(run_agent())