Implementation and Integration

85. An agent must maintain conversation context across turns without the application resending the full transcript. Which solution meets these requirements?

1. A Place the transcript in the agent's static instruction at the start of the conversation.
2. B Rely on the foundation model to retain the earlier turns between requests.
3. C Resend the complete transcript with every request regardless of length.
4. D Persist the session state keyed by a session identifier so context is retained between turns.

86. An agent occasionally loops, calling the same tool repeatedly without converging on an answer. Which combination of steps meets these requirements? (Select TWO.)

1. A Implement a circuit breaker that stops invoking a tool after repeated failures within a request.
2. B Increase the agent's request timeout so it has more opportunity to converge.
3. C Add more tools so the agent has additional paths to a solution.
4. D Increase the model's temperature so the agent varies its approach.
5. E Configure a maximum iteration limit per request with a defined fallback response.

87. A tool must be exposed to an agent with a standard interface so other agents can consume it without bespoke integration. Which solution meets these requirements?

1. A Expose the tool through a bespoke API per consuming agent.
2. B Implement the tool as a function inside each agent that needs it.
3. C Implement the tool behind a Model Context Protocol server so any compatible client can discover and invoke it.
4. D Document the tool's HTTP interface for each agent team to integrate against.

88. A workflow must pause for human review before an agent's proposed action is executed. Which solution meets these requirements?

1. A Orchestrate the workflow with AWS Step Functions and insert a task that waits for an approval callback before the action executes.
2. B Instruct the agent in its prompt to wait for approval before acting.
3. C Have the agent execute the action and notify a reviewer afterwards.
4. D Log the proposed actions for a reviewer to audit at the end of each day.

89. A complex request must be broken into reasoning steps with tool calls interleaved, and each step must be observable. Which solution meets these requirements?

1. A Invoke the model repeatedly from a script and log the final answer.
2. B Implement the reasoning and acting loop as an orchestrated workflow so each step and tool call is recorded as a discrete state.
3. C Implement the loop inside one Lambda function with the reasoning held in local variables.
4. D Ask the model to perform the whole reasoning chain in a single response.

90. A GenAI feature has unpredictable, spiky traffic and a strict cost constraint. Which deployment approach is appropriate?

1. A Deploy the model to a dedicated endpoint sized for the spike.
2. B Invoke the model on demand so cost follows usage, and monitor for throttling as volume grows.
3. C Purchase provisioned throughput sized for the largest expected spike.
4. D Purchase provisioned throughput sized for the average request rate.

91. A workload mixes short classification requests with long generation requests, and cost must be reduced without harming quality. Which solution meets these requirements?

1. A Route every request to the larger model for consistency.
2. B Route every request to the smaller model and accept the quality difference.
3. C Reduce the maximum output tokens for every request.
4. D Cascade the requests so a smaller model handles the classification and the larger model handles generation, validated by evaluation on each class.

92. A GenAI capability must be consumed by many internal applications with consistent authentication, throttling, and logging. Which solution meets these requirements?

1. A Have each application call the model service directly with its own credentials.
2. B Expose the capability through a central gateway layer that applies authentication, rate limiting, and logging before invoking the model.
3. C Publish a client library and ask each team to apply the controls consistently.
4. D Give each application its own copy of the prompts and guardrail configuration.

93. A GenAI component must be integrated into an existing order system without coupling the two release cycles. Which solution meets these requirements?

1. A Embed the GenAI component in the order system's codebase and deploy them together.
2. B Integrate through events so the order system publishes and the GenAI component consumes independently.
3. C Have the GenAI component poll the order database for new records.
4. D Call the GenAI component synchronously from within the order system's transaction.

94. A pipeline must deploy GenAI components with automated testing, security scanning, and the ability to roll back. Which solution meets these requirements?

1. A Deploy changes automatically on every commit without gating stages.
2. B Deploy changes to production and monitor for regressions afterwards.
3. C Build a CI/CD pipeline with automated evaluation and security scan stages that gate promotion, and deploy by reference so rollback is repointing.
4. D Deploy changes manually after a review meeting approves them.

95. A chat interface must display generated text as it is produced rather than after the full response completes. Which solution meets these requirements?

1. A Invoke the model synchronously and reduce the response length so it returns sooner.
2. B Invoke the model with a streaming API and deliver the chunks to the client over a streaming transport.
3. C Invoke the model synchronously and increase its maximum output tokens.
4. D Invoke the model asynchronously and poll for the completed response.

96. Model invocations intermittently fail with throttling during traffic spikes. Which combination of steps meets these requirements? (Select TWO.)

1. A Increase the maximum output tokens per request.
2. B Retry the failed calls immediately in a tight loop.
3. C Implement exponential backoff with jitter in the client.
4. D Disable the SDK's built-in retry behaviour so failures surface immediately.
5. E Raise available capacity through provisioned throughput or cross-Region inference.

97. Requests must be routed to different foundation models based on the content of each request. Which solution meets these requirements?

1. A Route every request to one model and accept the results.
2. B Invoke every candidate model and return the first response received.
3. C Ask the user to select the model before submitting the request.
4. D Implement content-based routing in an orchestration layer that selects the model per request from a configured mapping.

98. An engineering team must reduce the time spent writing boilerplate integration code for a GenAI application. Which solution meets these requirements?

1. A Copy the integration code from an existing application and adapt it by hand.
2. B Write the integration code from scratch to ensure it fits the use case.
3. C Use Amazon Q Developer to generate and refactor the integration code, reviewing its output before committing.
4. D Delay the integration work until a shared library is published.

99. A document processing workflow must extract, classify, and summarize incoming files with minimal custom code. Which solution meets these requirements?

1. A Process the documents manually and use the model only for summarization.
2. B Ask the foundation model to perform extraction, classification, and summarization from the raw file.
3. C Use Amazon Bedrock Data Automation to handle the extraction and processing stages within a managed workflow.
4. D Write Lambda functions that call each extraction and classification service in sequence.

100. An agent must decide between several specialised models depending on the subtask it is performing. Which approach is appropriate?

1. A Let the agent select the model itself at run time from its own judgement.
2. B Define a routing step that selects the model per subtask from measured capability on that subtask.
3. C Use the largest available model for every subtask.
4. D Use the cheapest available model for every subtask.

101. A multi-agent system must prevent two agents from making conflicting changes to the same resource. Which approach is appropriate?

1. A Run the agents sequentially so they never execute at the same time.
2. B Increase each agent's timeout so it can retry after a conflict.
3. C Serialize the changes through a single owning component or a lock, so only one change applies at a time.
4. D Instruct each agent in its prompt to check whether another agent is working.

102. An agent's tool must fail predictably when it receives a parameter it cannot process. Which approach is appropriate?

1. A Return an empty result so the agent continues.
2. B Validate the parameter against a schema and return a structured error the agent can act on.
3. C Retry the tool call with the same parameter.
4. D Return the tool's raw exception text to the agent.

103. A human review step must be inserted into an agentic workflow without blocking unrelated requests. Which approach is appropriate?

1. A Suspend the individual workflow execution awaiting an approval callback while other executions continue independently.
2. B Pause all agent executions until the reviewer responds.
3. C Have the agent poll for approval in a loop.
4. D Complete the action and ask the reviewer to reverse it if incorrect.

104. An agent must combine outputs from several models into one answer. Which consideration applies?

1. A Return whichever model responded first.
2. B Concatenate every model's output into the response.
3. C Return the longest of the model outputs.
4. D Define explicit aggregation logic, since combining model outputs requires a rule for resolving disagreement.

105. A generative workload must be deployed where GPU memory constrains how many concurrent requests each instance can serve. Which consideration applies?

1. A Concurrency per instance is bounded by the memory each in-flight request consumes, so capacity planning must account for both model and request memory.
2. B Concurrency is determined by the size of the container image.
3. C Concurrency can be raised indefinitely by increasing the request timeout.
4. D Concurrency is bounded only by the instance's vCPU count.

106. A cascading deployment must route routine queries away from the largest model. Which step establishes that the cascade is safe?

1. A Route a small percentage of all queries to the smaller model and observe the error rate.
2. B Compare the two models' published benchmark scores.
3. C Adopt the cascade because routine queries are simpler by definition.
4. D Evaluate the smaller model on a labelled sample of the routine query class and confirm it meets the quality bar.

107. A GenAI gateway must give the platform team visibility into which application made each model call. Which approach is appropriate?

1. A Record the model identifier used for each invocation.
2. B Require callers to authenticate to the gateway and record the caller identity with each invocation.
3. C Record the total number of invocations per hour.
4. D Ask each team to report its usage monthly.

108. A GenAI capability must be integrated with an on-premises system that cannot be exposed to the internet. Which approach is appropriate?

1. A Copy the on-premises data to AWS on a nightly schedule.
2. B Establish private connectivity between the AWS environment and the on-premises network, and invoke the model from the AWS side.
3. C Expose the on-premises system through a public endpoint restricted by IP address.
4. D Run the foundation model on the on-premises system.

109. A CI/CD pipeline for a GenAI component must prevent a prompt or model change from degrading quality in production. Which stage is required?

1. A An automated evaluation stage that compares the candidate against the current production baseline and blocks promotion on regression.
2. B A manual approval stage where a reviewer reads the change.
3. C A canary stage that routes a small percentage of traffic to the candidate.
4. D A rollback procedure documented in the runbook.

110. A client must handle a model response that is streamed incrementally and may end without completing. Which approach is appropriate?

1. A Retry the request whenever any chunk is delayed.
2. B Display whatever arrived and assume the response is complete.
3. C Buffer the whole stream before displaying anything.
4. D Detect the stream's completion signal explicitly and treat an incomplete stream as a failure rather than a partial success.

111. An application must observe latency across the retrieval, model invocation, and post-processing stages of each request. Which approach is appropriate?

1. A Instrument each stage as a trace segment so time is attributed to the stage that consumed it.
2. B Record the number of tokens generated per request.
3. C Record the total request duration and compare it against a target.
4. D Record the model's reported generation time.

112. An API fronting a foundation model must handle requests whose responses may exceed the client's expected size. Which approach is appropriate?

1. A Increase the client's timeout so large responses complete.
2. B Enforce a maximum output token limit and stream the response, so the client receives bounded, incremental output.
3. C Reject any request whose response might be large.
4. D Truncate the response at the API layer without informing the client.

113. A team must expose a GenAI capability to non-developers who need to compose their own workflows. Which approach is appropriate?

1. A Publish the model API documentation and an SDK example.
2. B Provide a visual workflow builder over prompt flows so users compose steps without writing code.
3. C Grant each user direct access to the model API.
4. D Provide a command line tool for invoking the model.

114. An engineer must diagnose why a GenAI application returns inconsistent answers to the same question. Which source should be examined first?

1. A The application's error logs for exceptions.
2. B The model provider's status page.
3. C The application's infrastructure metrics for the period.
4. D The logged prompts and responses for the affected requests, to establish whether the prompts differed.

115. An agent must persist its memory across sessions so a returning user is recognised. Which approach is appropriate?

1. A Rely on the model's context window.
2. B Store memory in the agent's process memory.
3. C Ask the user to re-explain their history each session.
4. D Store user-scoped memory in a durable store keyed by user identifier and load it into the agent's context at session start.

116. An agent's reasoning must be structured so each step is explicit and can be inspected. Which pattern applies?

1. A Select tools at random until one succeeds.
2. B Use a reason-then-act loop where the agent records its reasoning, chooses an action, observes the result, and repeats.
3. C Run a fixed script of tool calls regardless of input.
4. D Use a single prompt asking for the final answer.

117. An agent must be prevented from taking an irreversible action without confirmation. Which safeguard is appropriate?

1. A Limit the agent to read-only tools.
2. B Instruct the agent to be careful.
3. C Classify irreversible tools and require a human confirmation step before the agent invokes them.
4. D Log all actions for later review.

118. An MCP server must be deployed for a tool with a long-running, stateful operation. Which hosting choice is appropriate?

1. A A CloudFront distribution.
2. B A container on Amazon ECS, which supports long-running processes and persistent connections.
3. C A static file in Amazon S3.
4. D A Lambda function, which is limited in duration and stateless.

119. An MCP server must be deployed for a lightweight stateless tool invoked occasionally. Which hosting choice is appropriate?

1. A An ECS service with a minimum of two tasks.
2. B A database stored procedure.
3. C A Lambda function, which scales to zero and suits brief stateless invocations.
4. D A dedicated EC2 instance running continuously.

120. A multi-agent system must coordinate a supervisor agent delegating to specialist agents. Which orchestration pattern applies?

1. A A single agent handling everything.
2. B Agents communicating only through a shared file.
3. C All agents responding to every request independently.
4. D A supervisor that decomposes the task, dispatches subtasks to specialists, and aggregates their results.

121. A foundation model must be deployed for a workload where every invocation requires a GPU and traffic is continuous. Which deployment is appropriate?

1. A Deploy a SageMaker AI endpoint on GPU instances with auto scaling for the continuous load.
2. B Deploy a serverless inference endpoint.
3. C Bedrock on-demand invocation only.
4. D Lambda functions invoking the model.

122. A large language model must be loaded onto an instance whose GPU memory is smaller than the model's full precision size. Which technique applies?

1. A Increase the instance's CPU memory.
2. B Quantize the weights to lower precision so the model fits, validating quality impact.
3. C Reduce the context window.
4. D Load only half the model.

123. A GenAI feature must be embedded into a legacy application that exposes only a SOAP interface. Which integration is appropriate?

1. A Replace the legacy application.
2. B Build an adapter that translates between the legacy interface and the GenAI service's API, isolating each side from the other.
3. C Rewrite the legacy application to call the GenAI API directly.
4. D Have users copy data between the two systems manually.

124. A GenAI capability must be consumed by teams in a subsidiary whose data cannot leave its jurisdiction. Which architecture is appropriate?

1. A Deploy the capability in a Region within the jurisdiction, or on infrastructure within it, so processing occurs where the data may reside.
2. B Deploy centrally and encrypt the data in transit.
3. C Deploy centrally and delete the data after processing.
4. D Deploy centrally and have the subsidiary send its data.

125. A GenAI gateway must enforce a token budget per consuming application. Which implementation is appropriate?

1. A Set one budget for all applications combined.
2. B Review token usage monthly.
3. C Trust each application to stay within budget.
4. D Meter tokens per authenticated caller at the gateway and reject requests once the budget is exhausted.

126. A batch of documents must be processed by a foundation model without holding an HTTP connection open for each. Which approach is appropriate?

1. A Invoke the model synchronously for each document from a loop.
2. B Process the documents on a single thread.
3. C Open all connections in parallel and wait.
4. D Queue the documents with SQS and process them asynchronously from a worker, writing results to storage.

127. An API client must handle a streaming response over a transport that does not support server push. Which approach is appropriate?

1. A Use chunked transfer encoding or long polling to deliver the stream incrementally over the available transport.
2. B Reduce the response length.
3. C Switch to a different model.
4. D Wait for the full response and return it at once.

128. A static routing configuration sends all requests to one model, and the team wants to route by request type without redeploying. Which approach is appropriate?

1. A Hardcode the new rules and redeploy.
2. B Move the routing rules to a configuration store read at request time, so a rule change is a configuration update.
3. C Ask users to choose the model.
4. D Deploy a separate application per request type.

129. An API fronting a foundation model must handle a model timeout without leaving the client hanging. Which approach is appropriate?

1. A Set no timeout so the model always completes.
2. B Return an empty response on timeout.
3. C Set a timeout shorter than the client's, return a defined error or partial result, and retry with backoff where appropriate.
4. D Retry indefinitely until the model responds.

130. A team must build a customer-facing UI for a GenAI feature quickly with authentication and hosting handled. Which approach is appropriate?

1. A Have customers use the API without a UI.
2. B Use AWS Amplify to scaffold the UI with authentication and hosting integrated.
3. C Build the UI from scratch on EC2.
4. D Serve the UI from the model endpoint directly.

131. A CRM must be enhanced with a GenAI feature that summarizes customer interactions. Which integration is appropriate?

1. A Trigger a Lambda function from the CRM's event or webhook on interaction close, generate the summary, and write it back through the CRM's API.
2. B Replace the CRM with a GenAI application.
3. C Export CRM data nightly and summarize offline.
4. D Have agents copy interactions into a chat interface.

132. A team must diagnose a GenAI-specific error pattern across many requests. Which approach is appropriate?

1. A Wait for users to report the pattern.
2. B Read individual log entries by hand.
3. C Query the prompt and response logs with CloudWatch Logs Insights to find the pattern's frequency and the inputs that trigger it.
4. D Restart the application.

133. An agent must be prevented from looping indefinitely on a task it cannot complete. Which approach is appropriate?

1. A Allow the agent to continue until it succeeds.
2. B Increase the agent's timeout.
3. C Limit the number of reasoning steps and terminate with a defined outcome when the limit is reached.
4. D Restart the agent when it takes too long.

134. An agent's tool selection must be constrained to those appropriate for the current task. Which approach is appropriate?

1. A Order the tools by expected relevance.
2. B Expose every tool and validate the choice afterwards.
3. C Expose every tool and instruct the agent which to use.
4. D Expose only the tools relevant to the task in the agent's tool set for that invocation.

135. An agent must hand off a task to a specialist agent and receive the result. Which approach is appropriate?

1. A Have the specialist agent write its result to a shared file.
2. B Have both agents work on the task concurrently.
3. C Have the specialist agent respond directly to the user.
4. D Define the handoff as a structured request and response so the delegating agent receives a predictable result.

136. An agent's actions must be attributable to the user who initiated the task. Which approach is appropriate?

1. A Propagate the user's identity through the agent's execution and record it with each action.
2. B Record the identity only for actions that change data.
3. C Record the agent's own identity with each action.
4. D Record the actions without identity.

137. An agent must be evaluated before being given permission to act autonomously. Which approach is appropriate?

1. A Test it on a small number of examples.
2. B Review the agent's prompt and tool definitions.
3. C Run it in a mode where proposed actions are recorded but not executed, and review whether those actions would have been correct.
4. D Deploy it with full permissions and monitor the results.

138. A model endpoint must handle variable traffic without over-provisioning. Which approach is appropriate?

1. A Provision capacity for the peak.
2. B Provision the minimum capacity and accept throttling.
3. C Configure auto scaling on the endpoint with a target metric reflecting utilization.
4. D Invoke the model on demand without an endpoint.

139. A fine-tuned model must be deployed alongside the base model for comparison. Which approach is appropriate?

1. A Deploy the fine-tuned model to a separate application.
2. B Alternate between the models on a schedule.
3. C Deploy both behind a routing layer that sends a defined share of traffic to each and records the outcomes.
4. D Replace the base model with the fine-tuned model.

140. A model's deployment must be rolled back quickly if quality degrades. Which approach is appropriate?

1. A Keep the previous version deployed and shift traffic back to it on a quality signal.
2. B Retrain the previous version.
3. C Redeploy the previous version from its artifact.
4. D Disable the feature until the issue is resolved.

141. A GenAI capability must be exposed to several internal applications with consistent governance. Which approach is appropriate?

1. A Have each application call the model provider directly.
2. B Publish the model credentials to each team.
3. C Provide it through a shared internal service that applies guardrails, logging, and quota enforcement centrally.
4. D Deploy a separate model endpoint per application.

142. A GenAI application must integrate with an enterprise identity provider. Which approach is appropriate?

1. A Create local accounts within the application.
2. B Authenticate by network location.
3. C Use a shared service account for all users.
4. D Authenticate users through the provider and derive the application's authorization decisions from the resulting identity.

143. A GenAI feature must be added to an application without changing its existing release cycle. Which approach is appropriate?

1. A Add the feature to the application's codebase.
2. B Replace the application with a new one.
3. C Wait for the next major application release.
4. D Deploy the feature as a separate service the application calls, so it can be released independently.

144. Model invocations from several teams must be attributed for chargeback. Which approach is appropriate?

1. A Record the total invocations per day.
2. B Ask each team to report their usage.
3. C Divide the total cost equally between teams.
4. D Require each team to authenticate to the shared gateway and record token usage per authenticated caller.

145. A GenAI application must be deployed to several environments with environment-specific models. Which approach is appropriate?

1. A Hardcode the model and rebuild per environment.
2. B Maintain a separate codebase per environment.
3. C Parameterize the model identifier so the same deployment artifact serves every environment.
4. D Use the same model in every environment.

146. A GenAI service must enforce different rate limits for different consuming applications. Which approach is appropriate?

1. A Scale the backend to absorb all demand.
2. B Ask each team to limit their own usage.
3. C Apply a single limit across all callers.
4. D Identify each caller at the gateway and apply a per-caller quota.

147. An API client must handle a model response that exceeds the client's processing time budget. Which approach is appropriate?

1. A Wait for the complete response regardless of duration.
2. B Stream the response and process it incrementally, abandoning the request if the budget is exceeded.
3. C Increase the client's timeout indefinitely.
4. D Retry the request when it takes too long.

148. A foundation model API's throttling responses must be handled correctly. Which approach is appropriate?

1. A Back off exponentially with jitter and retry, distinguishing throttling from other errors.
2. B Retry immediately on any error.
3. C Increase the request rate to compensate.
4. D Treat throttling as a permanent failure.

149. An application must record the exact model version used for each response. Which approach is appropriate?

1. A Record the model configured at deployment.
2. B Record the model name without a version.
3. C Infer the version from the response's characteristics.
4. D Capture the model identifier and version from the API response and store it with the interaction record.

150. An application must handle a foundation model API returning a content filter refusal. Which approach is appropriate?

1. A Treat the refusal as a service error.
2. B Detect the refusal from the response and present an appropriate message rather than treating it as a failure.
3. C Strip the filter from the request.
4. D Retry the request unchanged.

151. A GenAI feature must be integrated into an existing web application's user interface. Which approach is appropriate?

1. A Return the complete response when generation finishes.
2. B Email the response to the user.
3. C Stream the response into the interface so the user sees progress, with a clear indication that the content is AI generated.
4. D Open the response in a separate window.

152. A GenAI application must let users correct an incorrect response. Which approach is appropriate?

1. A Ask the user to resubmit the request.
2. B Log the complaint without linking it to the interaction.
3. C Capture the correction linked to the interaction so it can inform evaluation and improvement.
4. D Escalate every correction to an engineer.

153. A GenAI capability must be made available to business users through a workflow tool. Which approach is appropriate?

1. A Expose the capability as a step the workflow tool can invoke, with its inputs and outputs defined.
2. B Provide direct model access to the business users.
3. C Have engineers build each workflow on request.
4. D Provide the API documentation to the business users.

154. An application must show users which sources a generated answer drew on. Which approach is appropriate?

1. A Provide a link to the whole knowledge base.
2. B Include the source names in the generated text.
3. C Return the retrieved source references alongside the answer and display them with the response.
4. D Log the sources without displaying them.

155. A GenAI application's responses must be consistent across the web and mobile clients. Which approach is appropriate?

1. A Implement the generation logic in each client.
2. B Use a different model for each client.
3. C Cache responses separately per client.
4. D Generate the response in a shared backend service that both clients call.

156. An application must handle a user request that requires information the model cannot access. Which approach is appropriate?

1. A Detect that the request falls outside the system's scope and say so, offering an alternative route where one exists.
2. B Retry with a larger model.
3. C Generate a plausible answer from the model's training data.
4. D Return an error without explanation.

157. A GenAI feature's rollout must be limited to a subset of users initially. Which approach is appropriate?

1. A Deploy to a separate environment users must opt into.
2. B Gate the feature behind a flag evaluated per user so exposure can be expanded gradually.
3. C Deploy during off-peak hours.
4. D Deploy the feature to all users and monitor.

158. A GenAI application's errors must be distinguishable from its refusals in operational metrics. Which approach is appropriate?

1. A Count refusals as successes.
2. B Emit distinct metrics for service errors and for policy refusals so each can be tracked separately.
3. C Record only the total request count.
4. D Count both as failures.

159. A GenAI application must degrade gracefully when its retrieval layer is unavailable. Which approach is appropriate?

1. A Return an error for every request.
2. B Retry retrieval until it succeeds.
3. C Respond without retrieval while clearly indicating that the answer is not grounded in the knowledge base, or decline where grounding is required.
4. D Generate from the model without indicating the limitation.