Workshop Explainer · Foundation

Amazon Bedrock

The managed platform for building generative AI applications and agents at production scale. This explainer walks the four pillars — model choice, your data (RAG), guardrails, and agents — then maps them to the integration and deployment decisions you'll make for a real agentic use case. Framed for a technical audience with basic AWS experience; sample scenarios use AnyCompany, a fictional bank.

☁️ Managed GenAI platform 🧠 Model choice 📚 Knowledge Bases / RAG 🛡️ Guardrails 🤖 AgentCore ⏱ ~20 min read

What Amazon Bedrock is

Bedrock is a fully managed, serverless service for building generative AI applications and agents. There are no servers to run and no models to host — you call an API, pick a model, and pay for what you use. Everything you need to go from prototype to production sits behind one unified API, secured by AWS IAM and kept inside your account.

💡The big picture: Bedrock has four pillars — model choice, your data (RAG), guardrails, and agents (AgentCore). Master how those four fit together and you can build almost any agentic use case. This explainer takes them one tab at a time.

The four pillars

🧠

Model choice

High-performing foundation models from Amazon (Nova), Anthropic (Claude), OpenAI, Meta, Mistral and more — all through one API. Swap models without changing application code.

📚

Your data (RAG)

Ground responses in your own content with Knowledge Bases — a fully managed RAG pipeline (ingest, retrieve, augment) with automatic citations.

🛡️

Guardrails

Configurable safeguards that filter harmful content, block prompt attacks, redact PII, and check groundedness — applied to any model, agent, or knowledge base.

🤖

Agents (AgentCore)

Build, deploy, and operate agents at scale with any framework or model — runtime, memory, identity, gateway, and observability as managed building blocks.

Why it matters now

📈

2023 → 2025

The market moved from proofs-of-concept (2023) to production (2024) to business value (2025). The questions shifted from "what is GenAI?" to "how do I lower cost, stay compliant, and ship agents?"

🏢

Enterprise adoption

Gartner projects 80%+ of enterprises will have used generative AI APIs or deployed GenAI-enabled apps by 2026.

🤖

Agents rising

Gartner projects 33% of enterprise apps will include agentic AI by 2028 (up from <1% in 2024), and 15% of day-to-day decisions made autonomously.

🧭The evolution to agentic AI: assistants (follow rules, single-turn) → agents (reason, break tasks into steps, call tools) → agentic systems (autonomous multi-step workflows, multi-agent). More capability means less human oversight per step — which raises the bar on guardrails, identity, and observability.

🧪 How to read this explainer

Work through the tabs in order — each maps to one pillar plus the deployment concerns around it:

Where this sits in the workshop. This explainer is the conceptual foundation. The hands-on build happens in the Strands Agents lab, where you wire a real agent to a Bedrock model, a Knowledge Base, and deploy it to AgentCore Runtime.
🛡️ Security first. Bedrock keeps your prompts, responses, and fine-tuned models inside your account and Region — they're never used to train the base models or shared with providers. When you build, give each workload least-privilege IAM (only the models and data it needs) and keep real customer data out of samples.

Model choice is the keystone

Bedrock gives you a broad selection of fully managed foundation models through one API. You can pick from leading providers, evaluate models against your own data, import your own customized models, and reach 100+ more via Bedrock Marketplace. Because the API is unified, you can swap the underlying model without rewriting your application.

🧠Four ways to get a model: choose from leading providers · evaluate to pick the right one · import your own customized model · access 100+ specialized models in Bedrock Marketplace.

Model families available

Amazon Nova Anthropic Claude OpenAI GPT & gpt-oss Meta Llama Mistral DeepSeek Cohere Command / Embed / Rerank AI21 Jamba Qwen3 Stability (Stable Image) Writer Palmyra + Bedrock Marketplace (100+)

Amazon Nova — frontier intelligence, leading price/performance

Nova Lite

Fast, cost-effective reasoning model for everyday tasks. Generally available.

🧠

Nova Pro

Nova's most intelligent model for complex AI workloads. Preview.

🎨

Nova Omni

All-in-one model for multimodal reasoning and image generation. Preview.

🔎

Nova Multimodal Embeddings

State-of-the-art embeddings for semantic search and agentic RAG. Generally available.

Choosing a Claude tier

Most agentic workloads land on Claude. The tradeoff is intelligence vs. cost vs. latency — pick the smallest model that clears your quality bar, then scale.

 Claude OpusClaude SonnetClaude Haiku
Best forHardest, most complex work needing peak intelligence & sustained autonomyHigh-volume production; near-Opus intelligence at scale; agentic workflowsLatency-sensitive UX; high-volume; executing tasks planned by bigger models
Cost (In/Out per MTok)$5.00 / $25.00$3.00 / $15.00$1.00 / $5.00
Relative latencyFast (~80 TPS)Fast (~80 TPS)Very fast (~109 TPS)
Context window1M (beta)1M (beta)200K
Max output128K tokens128K tokens64K tokens
OpenAI on Bedrock. Frontier GPT models and open-weight gpt-oss are available through the Bedrock API with native OpenAI SDK support — you can migrate with zero code changes.

Tune cost & latency without changing your model

🎚️

Service tiers

Priority, Standard, Flex, Batch, and Reserved — trade latency for cost depending on how time-sensitive the workload is.

🌏

Region options

In-Region for data-residency control, or Cross-Region inference for capacity and resilience (outputs still land in the home Region).

🚀

Feature-level

Prompt Caching (reuse repeated context), Intelligent Prompt Routing (send easy prompts to cheaper models), and Model Distillation.

Pick with evidence, not vibes

Bedrock Model Evaluation lets you compare models on your own data — automatic metrics (accuracy, robustness, toxicity), human review workflows, or LLM-as-a-judge for human-like quality at scale. Evaluate before you commit a model to production.
⌨️ Developer detail — bring or extend your own model

Custom Model Import lets you run your own fine-tuned/customized model behind the same serverless Bedrock API, on-demand, and still use native features like Knowledge Bases, Guardrails, and Agents. Bedrock Marketplace adds 100+ specialized models (from Hugging Face, Stability, Writer, NVIDIA, and others) through a serverful deployment when you need something niche.

Your data is the differentiator

A base model knows the public internet up to its training cut-off. It does not know your products, your policies, or last night's transactions. Retrieval-Augmented Generation (RAG) closes that gap: fetch relevant snippets from your data at query time and add them to the prompt, so the model answers from your knowledge — with citations. Bedrock Knowledge Bases is the fully managed way to do this.

RETRIEVE · FIND GROUNDED FACTS GENERATE · ANSWER FROM YOUR DATA 👤 User input UI / API call 1 🔢 Embed query [0.12 −0.03 0.95 …] 2 Semantic search over Knowledge Base 3 🧩 Augment prompt context + question 4 Generate Claude / Nova / GPT 5 💬 Cited answer with sources 6
👤1 · User input: A question enters through your app — a UI or an API call. The model won't answer yet; the orchestrator first goes to fetch grounded facts. Press Auto-Tour or click any step to follow the flow.
What Knowledge Bases handle for you. Ingestion (chunking + embedding), vector storage, semantic retrieval, prompt augmentation, multi-turn session context, and automatic citations — so answers are traceable and hallucinations are reduced. You don't build or operate the pipeline.

How much of the pipeline do you want to own?

RAG on Bedrock is a spectrum from full control to full abstraction. Pick based on how much you need to tune vs. how fast you want to ship.

🛠️

Build your own RAG

Full control over every component — choose your vector store, parsers, and connector workflow. You connect and retrieve data; you own the pipeline. Most control.

📚

Bedrock Knowledge Bases

Connect data, retrieve data — infrastructure and scaling handled for you. You still choose embedding model and vector store, but skip the plumbing.

Managed Knowledge Base

Connect data, users get answers — no RAG pipeline to build. Native AgentCore integration, agentic retrieval (retrieve → reason → re-rank), and native connectors. Most abstraction.

🔎Agentic retrieval goes beyond "closest vector match": it plans queries across multiple knowledge bases, connects concepts across documents, then evaluates and re-ranks candidates to return the best answer — not just the nearest chunk.

Bring your own vector store

OpenSearch Serverless Aurora PostgreSQL (pgvector) Amazon Neptune Amazon DynamoDB (zero-ETL) Amazon MemoryDB Pinecone MongoDB Redis Enterprise Cloud
Beyond plain text. Bedrock Data Automation extracts structured data from multi-modal content (documents, images, audio, video) through one API, and GraphRAG links entities across documents so answers can reason over relationships — not just isolated chunks. Native connectors reach SharePoint, Confluence, Google Drive, S3, and OneDrive with ACLs preserved.
🔗 Lab tie. In the Strands lab (Lab 4), the agent's built-in retrieve tool queries a Bedrock Knowledge Base exactly like step 2 above — this is the same RAG flow, wired into an agent.

Safety, privacy, and compliance

For a bank, the model is only usable if it's safe and the data stays protected. Bedrock addresses both: Guardrails control what goes in and comes out of the model, and the platform's data-handling guarantees keep your prompts and responses inside your account and Region.

Amazon Bedrock Guardrails

A configurable safety layer that works with any model, agent, or knowledge base — even models hosted outside Bedrock.

🚫

Content filters

Detect and filter up to 88% of undesirable or harmful text and image content, with thresholds you configure per category.

🛡️

Prompt-attack protection

Block jailbreaks, prompt injections, and prompt leakage before they reach the model.

🙈

PII redaction

Detect and redact personally identifiable and sensitive information in inputs and responses.

🧭

Denied topics

Define and disallow subjects with short natural-language descriptions — e.g. keep a support bot off investment advice.

Groundedness & relevance

Filter hallucinations by checking whether a response is grounded in the provided context and relevant to the query.

🔬

Automated Reasoning checks

Verify responses against encoded rules with up to 99% accuracy to minimize factual errors on policy-critical answers.

How Guardrails work — the architecture

the input is checked against your policies · if it passes, the model runs · the response is checked by the same policies · if it passes, the final answer is returned — otherwise it's blocked, or the PII is masked. Click any policy in the guardrail box to explore it.

Amazon Bedrock Guardrails — the same policies screen both the input and the output

👤 User User input ⚙️ Model inference ⚙️ Model response ② if input check passes 1 3 INPUT OUTPUT 🛡️ Guardrails The same policies — applied both ways Attach per-API-call to Converse / InvokeModel, or screen text anywhere with ApplyGuardrail 🚫Content filters Denied topics 🔒Sensitive info 🔤Word filters Grounding 🧮Auto reasoning 4 if output check passes ✅ Final response
Input vs. output, and what a block does: a failed input check stops the request before the model is ever called — no inference cost, and a safe blocked message is returned. A failed output check either blocks the response or masks it — e.g. redacting a card number to ••••4921 — before it reaches the user or the logs.

The policy that fired — details

Security questions to answer when you build

You are always in control of your data

🔒

Not used for training

Your prompts, responses, and fine-tuned models are never used to train the base models and are never shared with model providers.

🌏

Regional isolation

Customer model data is isolated per customer and stays in the AWS Region where it was created. (Cross-Region inference may process in a secondary Region, but outputs are stored in the home Region.)

🔑

Encryption & keys

Encrypted in transit (TLS 1.2+) and at rest (AES-256 via AWS KMS). Fine-tuned models can use your own KMS keys — only you can access them.

📜

Compliance

20+ standards including HIPAA eligibility, SOC 1/2/3, ISO, PCI, FedRAMP (Moderate & High), and GDPR. Monitor API activity with AWS CloudTrail.

🔐The one-liner for risk teams: if the data isn't stored, it can't leak, can't be seen by model vendors, and can't be reused by AWS. Bedrock only retains operational metrics (like billing usage), not your inference data.
Region note for UOB. Bedrock is available across many Regions including Asia Pacific (Singapore, Sydney, Tokyo, Seoul, Mumbai, Jakarta, and more). Some models are gated per Region — confirm the model you want is enabled in ap-southeast-1 before committing an architecture to it, or plan a Cross-Region inference profile.

From prototype to production

A working agent in a notebook is exciting — but getting it to production is where most projects stall. Agents run for long stretches, call many tools, need memory, need fine-grained identity, and need monitoring built for autonomous systems. Amazon Bedrock AgentCore is a set of modular managed services that supply exactly those pieces, so you don't rebuild the plumbing for every agent.

The chasm. Gartner projects over 40% of agentic AI projects will be cancelled by 2027 — the usual culprits are cost, unclear value, and insufficient security. The gap between an exciting POC and a trusted production agent is performance, security, governance, and scalability.

AgentCore — three value props

Time to value

Build powerful agents without the infrastructure and operational headaches — the runtime, memory, and gateway are managed.

🧩

Flexible

Use any framework and any model — Strands Agents, LangGraph, CrewAI, LlamaIndex, Google ADK, OpenAI Agents SDK. Protocols: MCP and A2A.

🔐

Trusted

Deploy secure, scalable, reliable agents your organization can trust — with identity, policy, and observability built in.

The building blocks

Mix and match — you don't have to adopt all of them at once.

GroupServiceWhat it gives your agent
Build
tools & memory
MemoryRetain context across turns and sessions; learn from past interactions; personalize.
GatewayTurn existing APIs and services into agent-callable tools (MCP-ready).
BrowserLet the agent browse the web in a managed, sandboxed browser.
Code InterpreterExecute code securely for calculation, data work, and file handling.
Deploy
securely at scale
RuntimeServerless runtime purpose-built for long-running, multi-modal agents (auto-scaling, isolated sessions).
IdentityFine-grained identity and access control — the agent reaches only the systems and data it should.
PolicyGovernance guardrails around what the agent is allowed to do.
Operate
gain insight
ObservabilityTrace and monitor agent behavior — the reasoning steps, tool calls, and outcomes.
EvaluationsMeasure agent quality over time so you can catch regressions.
🤝Where Strands fits: Strands Agents is a supported AgentCore framework. You build the agent locally with Strands, then deploy it to AgentCore Runtime as a secure, auto-scaling managed endpoint — exactly the arc of the hands-on lab (Lab 6).

Related agent services

🔀

Bedrock Flows

Visually design and accelerate GenAI workflows — chain prompts, models, knowledge bases, and logic.

🤖

Bedrock Agents

Managed multi-step task execution over your systems and data: pick a model, give instructions, attach data sources, specify actions.

Knowledge check

1. What are the four pillars of Amazon Bedrock?

Bedrock is organized around four pillars: a choice of foundation models, tools to ground responses in your data (RAG / Knowledge Bases), guardrails for safety, and AgentCore for building and running agents.

2. Your support bot must answer from AnyCompany's policy documents, with citations. What do you reach for?

RAG via a Knowledge Base fetches the relevant snippets at query time and returns citations — no retraining, and answers stay current as documents change.

3. Where does the customer's prompt and response data live?

Bedrock keeps your inference data inside your account and Region. It isn't used to train the base models and isn't shared with providers — only operational metrics (like billing) are retained.

Apply the four pillars

Here's how the pillars come together for a real agentic use case. The scenarios below use AnyCompany, a fictional bank — swap in your own workload and the architecture stays the same.

💬

Customer Support Agent

Uses: a model (Claude/Nova) + a Knowledge Base of product FAQs and policies + Guardrails to block off-topic or unsafe answers. Answers questions with citations, escalates when unsure.

🕵️

Fraud & Dispute Triage

Uses: an orchestrator agent routing to specialist sub-agents (fraud scoring, dispute evidence, compliance). AgentCore Identity scopes each agent's access; Observability traces every decision.

🏪

Merchant Ops Assistant

Uses: AgentCore Gateway to expose settlement, chargeback, and payout APIs as tools; Memory for multi-turn context; a Knowledge Base for merchant onboarding docs.

📄

Invoice / Document Agent

Uses: Bedrock Data Automation to extract structured fields from documents, a model to reason over them, and DynamoDB tools to record results.

✍️ Design activity — sketch your agent

Before the lab, sketch one agent for a workload you own. These five questions map directly onto the Bedrock building blocks:

Next step — build it. Take this design into the Strands Agents lab. You'll wire an agent to a Bedrock model, add tools, connect a Knowledge Base with retrieve, and deploy to AgentCore Runtime — the four pillars, hands-on.

🎉 What to take away

Sources: distilled from the Amazon Bedrock First Call Deck and AWS documentation. Sample scenarios use the fictional AnyCompany bank — no real UOB or customer data. Content rephrased for teaching.