Inference Providers

AWS Agent Platform is provider-agnostic by design. A single blueprint field — model.provider — selects the inference backend: Amazon Bedrock, the direct Anthropic API, any OpenAI-compatible endpoint via LiteLLM, or Google Vertex AI (Gemini). Switching providers requires no code changes — only a blueprint update.

The four providers

Value Backend Primary use case
bedrock Amazon Bedrock Converse API Default. AWS-native. GuardRails, FIPS, PrivateLink.
anthropic Anthropic API directly Direct Anthropic access without AWS mediation.
litellm Any OpenAI-compatible endpoint via LiteLLM Self-hosted proxy, vLLM, Ollama, OpenRouter, Cloudflare AI Gateway, or any custom router.
vertex Google Vertex AI (Gemini models) GCP-hosted Gemini models.

How provider dispatch works

The BlueprintLoader._build_model_config() method reads model.provider and constructs the appropriate Strands SDK model instance via a match/case dispatch. All provider imports are lazy — non-default provider packages are only imported when the blueprint actually selects them, so unused SDK packages have zero startup cost.

The dispatch is exhaustive: an unknown provider value raises BlueprintLoadError immediately at load time, before any inference request is made.

In this section

  • Overview — how to choose a provider and what each delivers
  • Amazon Bedrock — region setup, model IDs, IAM
  • Anthropic — direct API, api_key_env
  • LiteLLM — proxy configuration, headers, the custom_llm_provider mechanism, CVE safety pin
  • Google Vertex AI — GCP credentials, what is and is not forwarded
  • Structured Output — output_schema, Strands native forced-tool path, minimum SDK version

Table of contents