Implementation and integration

Agents, deployment and plumbing

Domain two is where a model stops being a demo and starts being a coworker with API access. Expect questions about agents, how models are deployed, and how GenAI plugs into systems that predate it.

Agents and tools

An agent is a loop: the model reasons, asks for a tool, your code runs it, and the result goes back in. With the Converse API that shows up as a toolConfig on the request, a stop reason of tool_use, and a toolResult you return in the next message. Amazon Bedrock AgentCore is the current managed path: Runtime hosts the agent, Gateway turns APIs and Lambda functions into MCP tools, Memory keeps short-term and long-term context, and Identity handles authentication. Amazon Bedrock Agents Classic, which wraps the loop with action groups (Lambda plus an OpenAPI schema or function definitions), knowledge bases and traces, is closed to new customers (opens in a new tab) but still in the exam guide. Open-source options such as Strands Agents and Agent Squad cover code-first and multi-agent designs, and the Model Context Protocol (MCP) standardises how agents discover tools. Small stateless MCP servers fit Lambda, while heavier tools suit containers on ECS.

Agents need guardrails of the plain kind: stopping conditions, timeouts, IAM boundaries and circuit breakers, because a loop with no exit is a billing event. Step Functions suits this well. Use a Retry and Catch for failures, a max-iterations counter for loops, and a callback with a task token when a human must approve a step.

Deployment

On-demand Bedrock invocation is the default and scales within a quota. Provisioned Throughput reserves model units for steady, predictable load. SageMaker AI endpoints host models you control, which also covers hybrid designs. Cheaper tiers help: route routine queries to a smaller model and escalate hard ones, a pattern called model cascading.

Enterprise integration

Wrap models behind API Gateway for validation, throttling and auth, hook them to events with EventBridge, and process slow jobs through SQS. Use least-privilege IAM per workload, federate identity from your enterprise IdP, and keep traffic private with VPC endpoints or hybrid connectivity. CI/CD (CodePipeline, CodeBuild) should test prompts and tools, scan for security issues and support rollback.

For users, stream tokens as they arrive (ConverseStream or InvokeModelWithResponseStream) over WebSockets or server-sent events. For reliability, retry with exponential backoff and jitter, then fall back. Never retry a validation error forever; it will be exactly as wrong the hundredth time.