Skip to content

Amazon Bedrock

Amazon Bedrock gives an agent access to foundation models from several providers through one AWS API. Strands connects to Bedrock through the BedrockModel class, which is the default model provider when you create an Agent without specifying one.

This page covers running an agent on Bedrock: configuring AWS credentials, selecting a model, and setting the options BedrockModel exposes. BedrockModel supports:

  • Text generation
  • Multimodal input (images, documents, and more)
  • Tool calling
  • Guardrails
  • Prompt caching for system prompts, tools, and messages
  1. AWS Account: You need an AWS account with access to Amazon Bedrock
  2. AWS Credentials: Configure AWS credentials with appropriate permissions

To use Amazon Bedrock with Strands, your IAM user or role needs the following permissions:

  • bedrock:InvokeModelWithResponseStream (for streaming mode)
  • bedrock:InvokeModel (for non-streaming mode)

Here’s a sample IAM policy that grants the necessary permissions:

{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": [
"bedrock:InvokeModelWithResponseStream",
"bedrock:InvokeModel"
],
"Resource": "*"
}
]
}

For production, scope the Resource down to specific model ARNs.

Strands uses boto3 (the AWS SDK for Python) to make calls to Amazon Bedrock. Boto3 has its own credential resolution system that determines which credentials to use when making requests to AWS.

For development environments, configure credentials using one of these methods:

Option 1: AWS CLI

Terminal window
aws configure

Option 2: Environment Variables

Terminal window
export AWS_ACCESS_KEY_ID=your_access_key
export AWS_SECRET_ACCESS_KEY=your_secret_key
export AWS_SESSION_TOKEN=your_session_token # If using temporary credentials
export AWS_REGION="us-west-2" # Used if a custom Boto3 Session is not provided

Option 3: Custom Boto3 Session

You can configure a custom boto3 Session and pass it to the BedrockModel:

import boto3
from strands.models import BedrockModel
# Create a custom boto3 session
session = boto3.Session(
aws_access_key_id='your_access_key',
aws_secret_access_key='your_secret_key',
aws_session_token='your_session_token', # If using temporary credentials
region_name='us-west-2',
profile_name='your-profile' # Optional: Use a specific profile
)
# Create a Bedrock model with the custom session
bedrock_model = BedrockModel(
model_id="global.anthropic.claude-sonnet-5-5",
boto_session=session
)

For complete details on credential configuration and resolution, see the boto3 credentials documentation.

Option 4: aws login

aws login provides browser-based authentication for temporary credentials. Requires AWS CLI version 2.32.0 or later.

Terminal window
aws login

To use aws login with enhanced performance, install botocore with CRT support:

Terminal window
pip install botocore[crt]

See the Login for AWS local development using console credentials documentation for more details.

A basic Agent uses the BedrockModel provider with Claude Sonnet 4.6 by default:

from strands import Agent
agent = Agent()
response = agent("Tell me about Amazon Bedrock.")

You can specify which Bedrock model to use by passing in the model ID string directly to the Agent constructor:

from strands import Agent
# Create an agent with a specific model by passing the model ID string
agent = Agent(model="global.anthropic.claude-sonnet-5-5")
response = agent("Tell me about Amazon Bedrock.")

Note: See Bedrock troubleshooting if you encounter any issues.

For more control over model configuration, you can create an instance of the BedrockModel class:

from strands import Agent
from strands.models import BedrockModel
# Create a Bedrock model instance
bedrock_model = BedrockModel(
model_id="us.amazon.nova-premier-v1:0",
temperature=0.3,
top_p=0.8,
)
# Create an agent using the BedrockModel instance
agent = Agent(model=bedrock_model)
# Use the agent
response = agent("Tell me about Amazon Bedrock.")

The BedrockModel supports various configuration parameters. For a complete list of available options, see the BedrockModel API reference.

Common configuration parameters include:

  • model_id - The Bedrock model identifier
  • temperature - Controls randomness (higher = more random)
  • max_tokens - Maximum number of tokens to generate; defaults to the model’s known output limit (get_config() shows the value); pass None to leave it unset. Bedrock reserves input + max_tokens against your tokens-per-minute quota when a request starts, so set it explicitly for high-concurrency workloads
  • streaming - Enable/disable streaming mode
  • guardrail_id - ID of the guardrail to apply
  • cache_prompt - Cache point type for the system prompt (deprecated, use cache_config)
  • cache_config - Configuration for prompt caching (e.g., CacheConfig(strategy="auto"))
  • cache_tools - Enable tool caching (deprecated, use cache_config with CacheConfig(tools_ttl=...))
  • strict_tools - Enforce structured output on tool definitions. Bedrock’s strict mode restricts which JSON Schema features tool input schemas may use (for example, oneOf is unsupported), so a schema that uses an unsupported feature fails at request time. See the Bedrock structured output documentation
  • boto_session - Custom boto3 session for AWS credentials
  • api_key - Amazon Bedrock API key for bearer token authentication; requests use the key instead of SigV4 signing
  • additional_request_fields - Additional model-specific parameters
from strands import Agent
from strands.models import BedrockModel
from botocore.config import Config as BotocoreConfig
# Create a boto client config with custom settings
boto_config = BotocoreConfig(
retries={"max_attempts": 3, "mode": "standard"},
connect_timeout=5,
read_timeout=60
)
# Create a configured Bedrock model
bedrock_model = BedrockModel(
model_id="global.anthropic.claude-sonnet-5-5",
region_name="us-east-1", # Specify a different region than the default
max_tokens=4096,
stop_sequences=["###", "END"],
boto_client_config=boto_config,
)
# Create an agent with the configured model
agent = Agent(model=bedrock_model)
# Use the agent
response = agent("Write a short story about an AI assistant.")

The TypeScript SDK applies a default requestTimeout of 120000 ms (120 seconds) when constructing the Bedrock Runtime client, since the underlying AWS SDK defaults to 0 (disabled), which lets a stuck connection hang. The timeout counts stream inactivity, so a long-thinking model or a large tool-call payload that keeps the stream quiet past the limit fails with Stream timed out because of no activity. Raise it with the requestTimeout option:

import { BedrockModel } from '@strands-agents/sdk/models/bedrock'
const bedrockModel = new BedrockModel({
modelId: 'global.anthropic.claude-sonnet-5-5',
requestTimeout: 600_000, // 10 minutes
})

The same value can also be set through clientConfig.requestHandler alongside other handler options; when both are given, the top-level requestTimeout wins:

const bedrockModel = new BedrockModel({
modelId: 'global.anthropic.claude-sonnet-5-5',
clientConfig: {
requestHandler: { requestTimeout: 600_000, connectionTimeout: 5_000 },
},
})

Passing a fully-constructed handler instance (rather than an options bag) bypasses both; the handler’s own timeouts apply unchanged and a requestTimeout option is ignored with a warning.

Some Bedrock models only support non-streaming tool use. Set the streaming configuration to false to use those models. Both modes give your agent the same event structure and behavior, since Strands converts non-streaming responses to the streaming format internally.

# Streaming model (default)
streaming_model = BedrockModel(
model_id="global.anthropic.claude-sonnet-5-5",
streaming=True, # This is the default
)
# Non-streaming model
non_streaming_model = BedrockModel(
model_id="us.meta.llama3-2-90b-instruct-v1:0",
streaming=False, # Disable streaming
)

See the Amazon Bedrock documentation for Supported models and model features to learn about the streaming support for different models.

Some Bedrock models accept multimodal input such as documents, images, and audio. Pass the content blocks directly to the agent:

from strands import Agent
from strands.models import BedrockModel
# Create a Bedrock model that supports multimodal inputs
bedrock_model = BedrockModel(
model_id="global.anthropic.claude-sonnet-5-5"
)
agent = Agent(model=bedrock_model)
# Send the multimodal message to the agent
response = agent(
[
{
"document": {
"format": "txt",
"name": "example",
"source": {
"bytes": b"Once upon a time..."
}
}
},
{
"text": "Tell me about the document."
}
]
)

For the full list of input types, see the API Reference.

As an alternative to providing media content as bytes, Amazon Bedrock supports referencing documents, images, videos, and audio stored in Amazon S3 directly. This is useful when working with large files or when your content is already stored in S3.

from strands import Agent
from strands.models import BedrockModel
agent = Agent(model=BedrockModel())
response = agent(
[
{
"document": {
"format": "pdf",
"name": "report.pdf",
"source": {
"location": {
"type": "s3",
"uri": "s3://my-bucket/documents/report.pdf",
"bucketOwner": "123456789012" # Optional: for cross-account access
}
}
}
},
{
"text": "Summarize this document."
}
]
)

Configure a guardrail on your BedrockModel to keep model output within your policies:

from strands import Agent
from strands.models import BedrockModel
# Using guardrails with BedrockModel
bedrock_model = BedrockModel(
model_id="global.anthropic.claude-sonnet-5-5",
guardrail_id="your-guardrail-id",
guardrail_version="DRAFT",
guardrail_trace="enabled", # Options: "enabled", "disabled", "enabled_full"
guardrail_stream_processing_mode="sync", # Options: "sync", "async"
guardrail_redact_input=True, # Default: True
guardrail_redact_input_message="Blocked Input!", # Default: [User input redacted.]
guardrail_redact_output=False, # Default: False
guardrail_redact_output_message="Blocked Output!", # Default: [Assistant output redacted.]
guardrail_latest_message=True, # Only evaluate the latest user message (default: False)
)
guardrail_agent = Agent(model=bedrock_model)
response = guardrail_agent("Can you tell me about the Strands Harness SDK?")

When a guardrail is triggered:

  • Input redaction (enabled by default): If a guardrail policy is triggered, the input is redacted
  • Output redaction (disabled by default): If a guardrail policy is triggered, the output is redacted
  • Custom redaction messages can be specified for both input and output redactions

Strands caches system prompts, tools, and messages on Bedrock to reduce token usage and latency. See Bedrock prompt caching for automatic and manual cache point placement, per-model token minimums, and cache metrics.

You can update the model configuration during runtime:

# Create the model with initial configuration
bedrock_model = BedrockModel(
model_id="global.anthropic.claude-haiku-5-5",
max_tokens=4096,
)
# Update configuration later
bedrock_model.update_config(
max_tokens=8192,
)

This is especially useful for tools that need to update the model’s configuration:

@tool
def update_model_id(model_id: str, agent: Agent) -> str:
"""
Update the model id of the agent
Args:
model_id: Bedrock model id to use.
"""
print(f"Updating model_id to {model_id}")
agent.model.update_config(model_id=model_id)
return f"Model updated to {model_id}"
@tool
def update_max_tokens(max_tokens: int, agent: Agent) -> str:
"""
Update the maximum number of tokens the agent's model can generate
Args:
max_tokens: Maximum number of tokens for the model to generate.
"""
print(f"Updating max_tokens to {max_tokens}")
agent.model.update_config(max_tokens=max_tokens)
return f"max_tokens updated to {max_tokens}"

Many models on Amazon Bedrock can reason through a problem before they answer. You turn reasoning on and set its depth through additional_request_fieldsadditionalRequestFields, which BedrockModel passes to the Converse API as additionalModelRequestFields. Each model family names these fields differently:

Model familyReasoning fieldsReference
Anthropic Claudethinking: {"type": "adaptive"}, with depth set by output_config.effort. Claude Opus 4.5, Claude Sonnet 4.5, Claude Haiku 4.5, and older models take thinking: {"type": "enabled", "budget_tokens": N} insteadAdaptive thinking, extended thinking
OpenAI GPTOpenAI Chat Completions fields, such as reasoning_effort: "high" (or "low" or "medium") on gpt-oss modelsOpenAI models

GPT models served only through the OpenAI-compatible endpoint, such as GPT-5.6, use the OpenAI Responses provider instead of BedrockModel. Set OpenAI’s reasoning parameter in its params; see OpenAI models through the OpenAI-compatible APIs. For other models, look up the fields on the model’s page in inference request parameters.

This example sets reasoning depth on Claude Opus 5.5:

from strands import Agent
from strands.models import BedrockModel
bedrock_model = BedrockModel(
model_id="global.anthropic.claude-opus-5-5",
additional_request_fields={
"thinking": {"type": "adaptive", "display": "summarized"},
"output_config": {"effort": "high"},
},
)
agent = Agent(model=bedrock_model)
response = agent(
"Two trains start 450 km apart and head toward each other at 120 km/h and 80 km/h. "
"A bird flies between them at 150 km/h until they meet. How far does the bird fly?"
)

Claude Opus 5.5 always reasons, at medium effort unless you set another level, and returns reasoning blocks with empty text unless you ask for display: "summarized".

Whichever model reasons, its reasoning arrives as reasoningContent blocks in the model response, separate from the response text. When you stream, Python callback handlers receive each reasoning delta with reasoning=True and reasoningText. In TypeScript, a ModelStreamUpdateEvent carries each reasoningContentDelta, and a ContentBlockEvent carries the assembled ReasoningBlock. See streaming events for the full event list.

Models can skip reasoning on simple requests, so don’t assume every response has a reasoning block. Read reasoning from these blocks rather than prompting the model to write it into its answer: newer Claude models, such as Claude Opus 5.5 and Claude Sonnet 5.5, may refuse requests that try to pull their reasoning into the response text.

Amazon Bedrock models support structured output through their tool calling capabilities. Pass a schema to the agent, and Strands converts it to Bedrock’s tool specification format and validates the response.

Define a Pydantic model and pass it to agent.structured_output():

from pydantic import BaseModel, Field
from strands import Agent
from strands.models import BedrockModel
from typing import List, Optional
class ProductAnalysis(BaseModel):
"""Analyze product information from text."""
name: str = Field(description="Product name")
category: str = Field(description="Product category")
price: float = Field(description="Price in USD")
features: List[str] = Field(description="Key product features")
rating: Optional[float] = Field(description="Customer rating 1-5", ge=1, le=5)
bedrock_model = BedrockModel()
agent = Agent(model=bedrock_model)
result = agent.structured_output(
ProductAnalysis,
"""
Analyze this product: The UltraBook Pro is a premium laptop computer
priced at $1,299. It features a 15-inch 4K display, 16GB RAM, 512GB SSD,
and 12-hour battery life. Customer reviews average 4.5 stars.
"""
)
print(f"Product: {result.name}")
print(f"Category: {result.category}")
print(f"Price: ${result.price}")
print(f"Features: {result.features}")
print(f"Rating: {result.rating}")

For schema patterns, error handling, and per-invocation overrides, see Structured Output.

Context management uses token counting to estimate input tokens before each model call.

The Bedrock provider can use the native count_tokens API via the CountTokens action in the Converse API. This includes system prompts, messages, and tool specifications in the count.

Not all Bedrock models support the CountTokens API. When a model doesn’t support it or the caller doesn’t have the required IAM permissions, the provider caches this result and falls back to estimation with a character-based heuristic (characters ÷ 4 for text, characters ÷ 2 for JSON) for subsequent calls.

You can enable native token counting with:

model = BedrockModel(
model_id="global.anthropic.claude-sonnet-5-5",
use_native_token_count=True,
)

To use a model that Amazon Bedrock only serves on bedrock-mantle, or a vendor feature that the Converse API doesn’t expose (such as an Anthropic beta), call the model through its vendor’s API instead of BedrockModel. For everything else, stay on BedrockModel: its Converse API gives every model on bedrock-runtime the same request shape.

bedrock-mantle isn’t a separate service or model catalog. It is Bedrock’s second endpoint family, and it serves Bedrock-hosted models through vendor APIs. The two families serve different, overlapping model sets: AWS lists which endpoint serves each model in endpoint availability by model.

To use a vendor API, build the vendor SDK’s client for Bedrock and pass it to the matching model provider as client. The vendor SDK signs requests with your AWS credentials, and the model provider runs the agent loop on top of it. We chose this over Bedrock-specific wrappers because the vendor SDKs already ship Bedrock clients and keep them current.

APIEndpointModel providerClient
Anthropic Messages (InvokeModel)bedrock-runtimeAnthropicModelAnthropicModelanthropic.AsyncAnthropicBedrockAnthropicBedrock
Anthropic Messagesbedrock-mantleAnthropicModelAnthropicModelanthropic.AsyncAnthropicBedrockMantleAnthropicBedrockMantle
OpenAI ResponsesbothOpenAIResponsesModelOpenAIModelopenai.AsyncOpenAIOpenAI
OpenAI Chat CompletionsbothOpenAIModelOpenAIModel with api: 'chat'openai.AsyncOpenAIOpenAI

The model provider doesn’t close a client you pass in. In Python, don’t share one client across asyncio event loops.

Pass Anthropic’s Bedrock client to AnthropicModel. It uses your AWS credentials from the standard credential chain.

import anthropic
from strands import Agent
from strands.models.anthropic import AnthropicModel
model = AnthropicModel(
model_id="global.anthropic.claude-sonnet-4-6",
max_tokens=4096,
client=anthropic.AsyncAnthropicBedrock(aws_region="us-west-2"),
)
agent = Agent(model=model)
response = agent("What is 2+2?")
print(response)

Requires pip install 'strands-agents[anthropic]'.

Two differences from the Anthropic API apply on bedrock-runtime:

  • Sampling parameters. With Python anthropic 2.x or TypeScript @anthropic-ai/sdk 1.x, Bedrock rejects temperature with Extra inputs are not permitted unless the request body pins the Bedrock API version. In Python, add "extra_body": {"anthropic_version": "bedrock-2023-05-31"} to params. In TypeScript, add anthropic_version: 'bedrock-2023-05-31' to params.
  • Throttling. In Python, a throttling error that arrives in the middle of an InvokeModel stream isn’t raised as ModelThrottledException, so the agent doesn’t retry it. Use BedrockModel if you rely on Strands retrying throttled requests.

To reach Claude on bedrock-mantle instead of InvokeModel on bedrock-runtime, swap in the bedrock-mantle client:

import anthropic
from strands import Agent
from strands.models.anthropic import AnthropicModel
model = AnthropicModel(
model_id="anthropic.claude-haiku-4-5",
max_tokens=4096,
client=anthropic.AsyncAnthropicBedrockMantle(aws_region="us-west-2"),
)
agent = Agent(model=model)
response = agent("What is 2+2?")
print(response)

AsyncAnthropicBedrockMantle requires anthropic 0.91 or later.

OpenAI models through the OpenAI-compatible APIs

Section titled “OpenAI models through the OpenAI-compatible APIs”

Point an OpenAI client at your region’s Bedrock endpoint and give it a function that mints a bearer token from your AWS credentials. The client calls that function on every request, so long-running agents survive token expiry.

import asyncio
from aws_bedrock_token_generator import provide_token
from openai import AsyncOpenAI
from strands import Agent
from strands.models.openai_responses import OpenAIResponsesModel
region = "us-east-1"
async def bedrock_token() -> str:
return await asyncio.to_thread(provide_token, region=region)
client = AsyncOpenAI(
base_url=f"https://bedrock-mantle.{region}.api.aws/v1",
api_key=bedrock_token,
)
model = OpenAIResponsesModel(model_id="openai.gpt-oss-120b", client=client)
agent = Agent(model=model)
response = agent("What is 2+2?")
print(response)

Requires pip install 'strands-agents[openai]' with openai 1.106 or later, the first release that accepts a function as api_key. For Chat Completions, pass the same client to OpenAIModel from strands.models.openai.

To authenticate with a Bedrock API key instead, pass the key string as the API key.

The base URL depends on the endpoint and, on bedrock-mantle, on the model:

EndpointBase URL
bedrock-runtimehttps://bedrock-runtime.{region}.amazonaws.com/openai/v1
bedrock-mantle, most modelshttps://bedrock-mantle.{region}.api.aws/v1
bedrock-mantle, newer lines such as openai.gpt-5.* and xai.grok-4.*https://bedrock-mantle.{region}.api.aws/openai/v1

A wrong base path fails with HTTP 400. Only bedrock-runtime accepts cross-Region inference profile IDs such as global.openai.gpt-5.6-luna. AWS compares the two endpoints in endpoints supported by Amazon Bedrock.

To skip building the client yourself, pass bedrock_mantle_configbedrockMantleConfig. The provider derives the endpoint from your region, picks the bedrock-mantle base path for the model, and mints bearer tokens from your AWS credentials:

from strands import Agent
from strands.models.openai_responses import OpenAIResponsesModel
model = OpenAIResponsesModel(
model_id="openai.gpt-oss-120b",
bedrock_mantle_config={"region": "us-east-1"},
)
agent = Agent(model=model)
response = agent("What is 2+2?")
print(response)

Omit the region to resolve it from your AWS environment. The config also accepts AWS credentials to forward to the token generator (credentials_providercredentials, a botocore CredentialProviderstatic identity or provider function) and a token-lifetime override (expiryexpiresInSeconds). In Python, boto_session can also supply the region from a configured profile, and "endpoint": "bedrock-runtime" targets bedrock-runtime instead of bedrock-mantle.

If you encounter the error:

Invocation of model ID XXXX with on-demand throughput isn’t supported. Retry your request with the ID or ARN of an inference profile that contains this model.

This typically indicates that the model requires Cross-Region Inference, as documented in the Amazon Bedrock documentation on inference profiles. To resolve this issue, prefix your model ID with the appropriate regional identifier (us.or eu.) based on where your agent is running. For example:

Instead of:

anthropic.claude-sonnet-5-5

Use:

us.anthropic.claude-sonnet-5-5

If you encounter the error:

ValidationException: An error occurred (ValidationException) when calling the ConverseStream operation: The provided model identifier is invalid

This is very likely due to calling Bedrock with an inference model id, such as: us.anthropic.claude-sonnet-5-5 from a region that does not support inference profiles. If so, pass in a valid model id, as follows:

agent = Agent(model="anthropic.claude-3-5-sonnet-20241022-v2:0")

CacheConfig with ARN-based inference profiles

Section titled “CacheConfig with ARN-based inference profiles”

If you’re using an ARN-based application inference profile as your model_id (e.g., arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123), CacheConfig(strategy="auto") will not automatically enable prompt caching.

The strategy="auto" detection checks the model ID string for "claude" or "anthropic" substrings. Cross-region inference profile IDs like us.anthropic.claude-sonnet-5-5 contain "anthropic" and are detected correctly, but application inference profiles use opaque resource IDs (application-inference-profile/abc123) that carry no model name information, so detection returns None, caching is skipped, and Strands logs a warning: model_id=<your-arn> | cache_config is enabled but this model does not support automatic caching.

Use strategy="anthropic" explicitly to fix this:

from strands import Agent
from strands.models import BedrockModel, CacheConfig
bedrock_model = BedrockModel(
model_id="arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123",
cache_config=CacheConfig(strategy="anthropic")
)
agent = Agent(model=bedrock_model)

strategy="anthropic" has identical performance to strategy="auto" and requires no additional API calls or IAM permissions.

Thinking block is bound to a different conversation

Section titled “Thinking block is bound to a different conversation”

When Amazon Bedrock rejects a request with this error:

ValidationException: The block is bound to a different conversation.

The provider retries the request once and logs this warning:

model_id=<...> | history edit invalidated replayed thinking blocks, retrying with drop_block

Claude models that preserve thinking bind each thinking block to the conversation that came before it. Summarizing or removing messages with context management or a conversation manager, restoring a session, or editing messages from a hook changes that earlier conversation, so the thinking blocks that remain in the history no longer match it. Accounts that enforce this binding, the default for accounts created on or after August 31, 2026, reject the next request.

The retry sets prefix_mismatch_behavior to drop_block in the thinking.block_binding request field, with "type": "adaptive" when no thinking is configured, and adds the thinking-binding-controls-2026-08-01 beta to any betas already in anthropic_beta. Claude then skips the stale thinking blocks instead of rejecting the request. The stale blocks stay in the history, so the provider keeps sending drop_block for that model ID until the model instance is discarded. Requests before the first rejection are sent unchanged.

  • Your block_binding wins. If the thinking configuration already contains block_binding, the provider sends it as written and does not retry. Set it yourself to choose the behavior up front or to keep the error.
  • Dropping has a cost. Claude loses the reasoning in the dropped blocks, the prompt cache restarts at the edited message, and responses can use more output tokens.
  • Some models require adaptive thinking. Claude Sonnet 5.5 and Claude Haiku 5.5 accept drop_block only with "type": "adaptive". With any other thinking type, the retry fails and its error is raised.

BedrockModel does not support server-side summary, so a message-level summary from context management is always client-side and rewrites that earlier conversation. To keep thinking, see Models that preserve thinking.