Amazon Bedrock

provider: bedrock is the default inference provider. It routes model calls through the Amazon Bedrock Converse API using the Strands SDK BedrockModel. No API key is needed — authentication is handled entirely through IAM.

Required environment variable

Variable	Purpose
BEDROCK_REGION	AWS region where Bedrock model requests are sent. Required. The loader raises BlueprintLoadError at startup if this variable is absent.

There is no region field in the model: block. The Bedrock region must be set as an environment variable. This is a deliberate separation: the blueprint declares what model to use; the deployment environment declares where it is accessed.

Blueprint configuration

model:
  provider: bedrock
  model_id: us.anthropic.claude-sonnet-4-20250514-v1:0
  temperature: 0.3
  max_tokens: 4096

All four fields are required. provider defaults to bedrock, so you may omit it:

model:
  model_id: us.anthropic.claude-sonnet-4-20250514-v1:0
  temperature: 0.1
  max_tokens: 8192

Runtime-switchable model IDs

Use ${VAR:-default} expansion to select the model at deploy time without changing the blueprint file:

model:
  provider: bedrock
  model_id: ${BEDROCK_MODEL_ID:-us.anthropic.claude-sonnet-4-20250514-v1:0}
  temperature: 0.1
  max_tokens: 8192

Set BEDROCK_MODEL_ID in the Runtime environment variables (via the agents Terraform module) to override the default at deploy time.

Model IDs

Bedrock model IDs are region-qualified. Cross-region inference profiles use the us. or eu. prefix:

# US cross-region inference profile
us.anthropic.claude-sonnet-4-20250514-v1:0
us.anthropic.claude-haiku-4-20250514-v1:0

# Direct regional model
anthropic.claude-3-5-sonnet-20241022-v2:0

# Amazon Nova
us.amazon.nova-pro-v1:0

Refer to the Amazon Bedrock model IDs documentation for the full current list. Model availability varies by region — verify the model is available in your BEDROCK_REGION before deploying.

IAM requirements

The agent’s IAM role (provisioned by modules/agents/iam.tf) needs bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream on the target model ARN, plus cross-region inference profile ARNs when using us.*/eu.* prefixed IDs.

The modules/platform/modules/agentcore/ sub-module provisions the Gateway IAM role with the required Bedrock permissions. The per-agent role in modules/agents/iam.tf holds the runtime-level model invocation rights.

Bedrock Guardrails

When data_protection.provider: bedrock is set in the observability block, the platform registers a GuardrailHook that calls Bedrock’s ApplyGuardrail API on agent I/O. This guard only activates when both conditions are met:

1. model.provider: bedrock
2. BEDROCK_GUARDRAIL_ID environment variable is set

If either condition is absent, the Bedrock guardrail hook is silently skipped — no error is raised. For non-Bedrock providers, use data_protection.provider: presidio for provider-agnostic in-process PII filtering. See Data Protection for the full comparison.

observability:
  data_protection:
    provider: bedrock   # Only activates when model.provider is also bedrock
    # BEDROCK_GUARDRAIL_ID and BEDROCK_GUARDRAIL_VERSION must be set as env vars
    # guardrail_id_env: BEDROCK_GUARDRAIL_ID        # Override env var name if needed
    # guardrail_version_env: BEDROCK_GUARDRAIL_VERSION

Extended thinking

The thinking: top-level block in the blueprint enables Claude’s extended thinking mode on Bedrock:

model:
  provider: bedrock
  model_id: us.anthropic.claude-sonnet-4-20250514-v1:0
  temperature: 0.1
  max_tokens: 16384

thinking:
  enabled: true
  budget_tokens: 10000   # Tokens reserved for the thinking phase

Extended thinking is currently only meaningful with Claude models on Bedrock. The budget_tokens value must be less than max_tokens.

Prompt caching

Bedrock supports prompt caching. The cache_prompt and cache_tools fields control the caching policy:

model:
  provider: bedrock
  model_id: us.anthropic.claude-sonnet-4-20250514-v1:0
  temperature: 0.1
  max_tokens: 4096
  cache_prompt: default   # default | none | <custom-cache-key>
  cache_tools: default

These fields are passed through to the Strands BedrockModel. Refer to Amazon Bedrock prompt caching documentation for model-specific caching support and pricing.

---

Related: Structured Output — output_schema works with Bedrock using the Strands native forced-tool path.