Amazon Bedrock

Learning Objectives

• Understand what Amazon Bedrock is and how its multi-model approach differs from single-vendor AI platforms
• Identify Bedrock's key capabilities: Knowledge Bases (RAG), Agents, Guardrails, and model evaluation
• Evaluate when Bedrock's managed platform is the right choice versus direct API access or competing cloud AI platforms

What Is Amazon Bedrock?

Amazon Bedrock is AWS's fully managed AI platform that provides access to multiple foundation models from different providers — all through a single, unified API. Instead of signing up for separate accounts with OpenAI, Anthropic, Meta, and Mistral, Bedrock lets you access Claude, GPT-5.5 (added April 28, 2026), Llama, Mistral Large, Cohere Command, Amazon Nova, and other models from one service within your existing AWS account. The April 2026 OpenAI partnership was a structural shift: it was the first time OpenAI's frontier models, the Codex coding agent, and a new Bedrock Managed Agents capability shipped through any cloud outside Microsoft Azure since 2019.

But Bedrock is more than a model gateway. It includes built-in capabilities for building production AI applications: Knowledge Bases for RAG (connecting models to your data), Agents for multi-step autonomous workflows, Guardrails for content filtering and safety, and Model Evaluation for comparing model performance on your specific tasks. These managed features eliminate significant engineering work that teams would otherwise build from scratch.

The core value proposition is straightforward: use the best model for each task, switch between providers without code changes, and keep everything within AWS's security and compliance framework — all on a single bill.

Pricing

All pricing is usage-based with no minimum commitment. Costs appear on your standard AWS bill alongside other services.

Core Capabilities

Multi-Model Access

Bedrock's defining feature is access to models from multiple providers through one API:

• Anthropic: Claude Opus, Sonnet, Haiku — strong reasoning and analysis (Amazon has invested $8 billion in Anthropic)
• OpenAI: GPT-5.5, GPT-5.4, plus the Codex coding agent and Bedrock Managed Agents (limited preview, added April 28, 2026) — the first non-Azure cloud distribution of OpenAI's frontier models since 2019
• Meta: Llama 3 and 4 variants — open-weight models for cost-sensitive workloads
• Mistral: Mistral Large 2, Mistral Small — European AI with multilingual strength
• Cohere: Command R+ — optimized for RAG and enterprise search
• Amazon: Nova Premier, Lite, Micro — native AWS models with tiered pricing
• Stability AI: Stable Diffusion — image generation
• AI21 Labs: Jamba — long-context language models

Switching between models requires changing only the model ID in your API call — no code restructuring, no new SDKs, no additional authentication.

Knowledge Bases (Managed RAG)

Bedrock Knowledge Bases provide fully managed Retrieval-Augmented Generation:

1. Connect data sources — point to S3 buckets, web crawlers, or Confluence/SharePoint
2. Automatic ingestion — Bedrock chunks documents, generates embeddings, and stores them in a managed vector store (or your own OpenSearch/Pinecone)
3. Query with RAG — ask questions and Bedrock retrieves relevant chunks, injects them into the model prompt, and returns grounded responses with source citations

This eliminates the need to build and maintain a custom RAG pipeline with separate embedding models, vector databases, and retrieval logic.

Bedrock AgentCore — Stateful Runtime

Bedrock AgentCore is the latest evolution of Bedrock's agent capabilities, introducing a stateful runtime with memory streaming. Unlike earlier agent implementations that were stateless between invocations, AgentCore maintains persistent state across agent sessions — enabling more complex, long-running workflows where agents can remember prior context and resume tasks. This is complemented by Amazon's Cerebras CS-3 partnership, which provides ultra-fast inference for latency-sensitive agent workloads.

Agents for Multi-Step Workflows

Bedrock Agents execute autonomous multi-step tasks by combining foundation models with your business APIs:

• Define actions by connecting to Lambda functions, APIs, or knowledge bases
• The agent plans a sequence of steps, executes them, handles errors, and returns results
• Built-in conversation memory across turns
• No orchestration framework required — Bedrock manages the agent loop

Guardrails

Configurable content filtering applied to model inputs and outputs:

• Content filters: Block harmful, offensive, or off-topic content by category and severity
• Denied topics: Prevent the model from engaging with specific subjects (e.g., competitor products, legal advice)
• Word filters: Block specific terms or patterns
• PII detection: Automatically detect and redact personally identifiable information
• Grounding checks: Validate that model responses are supported by the provided source documents

Guardrails are applied as a layer on top of any Bedrock model — configure once and apply across all your applications.

Strengths

• Multi-model flexibility: Access Claude, Llama, Mistral, Cohere, Nova, and more through one API — switch models without code changes
• Managed RAG pipeline: Knowledge Bases handle document ingestion, embedding, storage, and retrieval — no custom infrastructure
• Enterprise security: IAM, VPC endpoints, encryption, CloudTrail auditing — inherits your existing AWS security posture
• Guardrails built-in: Content filtering, PII detection, and grounding checks as a managed service — not a DIY implementation
• Agent orchestration: Multi-step autonomous workflows without a separate orchestration framework
• Unified billing: All model usage, RAG queries, and agent executions on your AWS bill — no separate vendor invoices

Limitations & Considerations

• AWS lock-in: Bedrock-specific features (Knowledge Bases, Agents, Guardrails) are not portable to other clouds — migration requires rebuilding
• Model availability lag: New model releases sometimes appear on Bedrock days or weeks after they are available via direct provider APIs
• Complexity: The breadth of features (models, Knowledge Bases, Agents, Guardrails, evaluation) creates a steep learning curve for teams new to AWS
• Indirect provider relationship: Issues with a specific model (e.g., Claude) are handled through AWS support, not directly with the model provider

Best Use Cases

When to choose alternatives:

• Running on Azure → Azure AI Foundry provides equivalent multi-model access with Azure integration
• Running on Google Cloud → Vertex AI with Gemini and Model Garden
• Want direct model provider access → Use Anthropic, OpenAI, or Mistral APIs directly for latest features
• Need open-source orchestration → LangChain or LlamaIndex for vendor-neutral RAG and agent frameworks

Getting Started

1. Open the Bedrock console — navigate to Amazon Bedrock in your AWS account and select your preferred region
2. Enable model access — request access to the models you want to use (Claude, Llama, Nova, etc.) from the Model Access page
3. Test in the Playground — compare models side-by-side using the same prompts in the Bedrock Playground
4. Create a Knowledge Base — point to an S3 bucket with your documents and let Bedrock build a managed RAG pipeline
5. Make API calls — use the Bedrock Runtime API via boto3 to invoke models programmatically with invoke_model
6. Add Guardrails — create a guardrail configuration and apply it to your model invocations for content safety

Key Takeaways

• Amazon Bedrock is AWS's managed multi-model AI platform — access Claude, GPT-5.5, Llama, Mistral, Nova, Cohere, and more through a single API within your AWS account
• The April 28, 2026 OpenAI partnership broke a seven-year Microsoft Azure cloud-distribution lock — GPT-5.5, GPT-5.4, Codex, and Bedrock Managed Agents are now in limited preview on AWS
• Built-in Knowledge Bases provide managed RAG (document ingestion, embedding, retrieval) without building custom vector database infrastructure
• Agents orchestrate multi-step workflows and Guardrails enforce content safety — both as managed services, not custom code
• Best for enterprises already on AWS that want multi-model flexibility with enterprise security; compare with Azure AI Foundry (Azure) and Vertex AI (Google Cloud)