Implementation and Integration26% · 5 topics

Q1. Which is the most appropriate description of how a ReAct-style agent works?

1. Regardless of the question, call every tool once in a fixed order and return the concatenated results
2. Generate an answer only once at the start and return it to the user without using tools
3. Have the user choose which tool to use every time and act exactly on that instruction
4. Repeat think → use a tool → observe the result, and answer once the goal is reached

Q2. You are defining tools for an agent. What is most important so the model chooses tools correctly?

1. Provide as many tools with similar names and functions as possible and let the model choose
2. Do not define tool input formats and accept any string
3. Make tool names and descriptions clear and show argument meaning and types in an input JSON Schema
4. Name tools with short numbers only and let the model guess without descriptions

Q1. Which statement about Bedrock Provisioned Throughput is incorrect?

1. Only tokens used are billed, with no charge for idle time
2. You purchase model units to secure a fixed throughput
3. Longer commitment terms such as 1 or 6 months give discounted hourly rates
4. It may be usable for inference with custom models

Q2. Throttling on on-demand Bedrock increases at peak times, and you are considering buying Provisioned Throughput. What is the most appropriate next step?

1. Without investigating the cause, just greatly increase app retry counts
2. Without measuring needs, buy as many model units as possible for the longest term
3. Ban all use at peak times and tell users to use it late at night or early morning
4. Measure tokens per minute at peak in CloudWatch and estimate the number of model units needed

Q1. Most steps are fixed, but you want to use FM judgment (summarization or classification) at some stages, with reliable control of the overall flow. Which architecture is most appropriate?

1. Instruct a Bedrock agent on the steps and let it decide tool call order
2. Chain Lambda functions so each calls the next one directly
3. Define the flow in Step Functions and use the Bedrock invoke integration where needed
4. Process every step with FM nodes in Bedrock Prompt Flows

Q2. You want an on-premises workflow system to call the assistant API without going over the internet. Which architecture is most appropriate?

1. Call an API Gateway private API via a VPC endpoint over Direct Connect or VPN
2. Attach an IP-restricting resource policy to an API Gateway Regional API
3. Put the API behind CloudFront and restrict access with AWS WAF
4. Create a Lambda function URL and restrict callers with IAM auth

Q1. An agent's tool calls an external API. You want the agent to stay stable even when the API is slow or fails. Which implementation is most appropriate?

1. Set timeouts and retries, return failures to the model as structured errors, and let it decide the next action
2. Repeat the same call any number of times with no wait or timeout until the API succeeds
3. When the API fails, hide the error and return a “succeeded” result to the model
4. Ignore all API failures and immediately end the whole agent process

Q2. You want every AWS account in the company to use only specific approved models. Which approach is most appropriate?

1. Grant permissions to each team's role with IAM Identity Center permission sets
2. Enable only approved models in each account's Bedrock model access
3. Use an Organizations SCP to deny InvokeModel except on approved model ARNs
4. Detect and notify unapproved model use with AWS Config rules

Q1. You want Bedrock Agents actions to run inside your own app rather than Lambda (to call systems inside the internal network). Which approach is most appropriate?

1. Expose internal systems to the internet so the agent can call them freely
2. Use return of control so the agent returns the chosen operation and arguments to the app to execute
3. Stop using actions and have the agent guess internal system information
4. Write all action logic in the prompt and let the model operate internal systems directly without permission limits

Q2. You want to add an “order note summary” feature to an existing order app without changing the app's response speed. Which integration method is most appropriate?

1. Make summarization a separate app and have staff copy and paste order content by hand each time
2. Run summarization (Lambda, etc.) asynchronously on the order-saved event and display the summary on screen later
3. Call the FM synchronously inside the order-save process and make saving wait until the summary is ready
4. Do not save order notes and have the FM imagine and summarize the content each time