Section 3: Developing custom agents

Question 1, 3.2 Integrating enterprise domain knowledge

A retailer's ADK agent must look up orders in a Cloud SQL for PostgreSQL database. Security forbids the model from generating arbitrary SQL. Only three approved parameterized queries may run. The team also wants connection pooling and OpenTelemetry tracing, and does not want to write or maintain a custom MCP server. What should you do?

1. Run MCP Toolbox for Databases with the prebuilt PostgreSQL configuration, and have the agent call its generic execute_sql tool for the three approved queries.
2. In MCP Toolbox for Databases, define a Cloud SQL source and three parameterized SQL tools in tools.yaml, group them in a toolset, and load it into the agent.
3. Build a FastMCP server on Cloud Run that implements the three parameterized queries, a connection pool and OpenTelemetry tracing, and connect it as a toolset.
4. Grant the agent's identity the Cloud SQL Client role, and use the system instructions to tell it to run only the three approved SELECT statements.

Question 2, 3.1 Designing and building agentic workflows in code

An insurance company wants its claims-intake agent, built with Agent Development Kit (ADK), to use an open-weight Llama model for summarizing adjuster notes. Traffic is unpredictable and bursty, and it spikes after storms and is near zero otherwise. The team is three application developers with no MLOps or GPU capacity-planning experience, and leadership wants the agent in production within two weeks. The standard instruction-tuned model meets quality requirements. Which approach meets these requirements with the LEAST operational overhead?

1. Deploy the model from Model Garden to a dedicated Agent Platform endpoint on NVIDIA L4 GPUs and pass the endpoint resource name to the LlmAgent model parameter.
2. Call the model through Model as a Service (MaaS) on Agent Platform by using the LiteLlm wrapper as the LlmAgent model.
3. Serve the model on a GKE Autopilot cluster with vLLM and point the agent at the OpenAI-compatible endpoint through LiteLlm.
4. Build a custom vLLM container image, deploy it to an Agent Platform endpoint, and add autoscaling rules for the storm season.

Question 3, 3.1 Designing and building agentic workflows in code

A legal-tech company has standardized on Anthropic Claude for contract analysis because of internal benchmark results. It is now building an ADK agent that will run on Google Cloud. The security team forbids storing third-party API keys in application configuration and requires that all model calls be authorized with Google Cloud IAM through Application Default Credentials and billed to the Google Cloud account. What should the developer do?

1. Pass the Claude on Agent Platform model string to LlmAgent, set GOOGLE_GENAI_USE_ENTERPRISE=TRUE, and install anthropic[vertex].
2. Wrap Anthropic's public API with the LiteLlm connector and store the Anthropic API key in the project's .env file, which is excluded from Git.
3. Download the Claude weights from Model Garden and self-deploy them to a dedicated GPU endpoint that ADK calls by its endpoint resource name.
4. Replace Claude with a Gemma model that is fine-tuned to mimic Claude's contract-analysis outputs.

Question 4, 3.1 Designing and building agentic workflows in code

A meal-planning agent built with ADK uses Agent Platform Memory Bank through VertexAiMemoryBankService. Memories about each user's allergies and dietary restrictions are generated from past conversations. In testing, the agent sometimes suggests recipes that contain allergens because the model does not always decide to look up memories before it answers. Product owners require that the user's known restrictions be in the model's context on every turn. Which change should the developer make?

1. Replace the existing LoadMemoryTool with a custom tool that calls tool_context.search_memory, and include a stronger docstring.
2. Add callback_context.add_session_to_memory to the after_agent_callback so that memories are regenerated after every turn.
3. Add PreloadMemoryTool to the agent so that memories are retrieved at the start of each turn and included in the system instruction.
4. Switch to InMemoryMemoryService so that memory retrieval does not depend on a remote service call.

Question 5, 3.1 Designing and building agentic workflows in code

A bank's advisory agent generates memories through GenerateMemories calls that ADK callbacks trigger, and it occasionally writes memories directly with CreateMemory. A new retention policy requires every memory to be deleted automatically 90 days after it was last created or updated, with no custom cleanup code. Memory revisions are already covered by a separate policy. How should the architect configure Agent Platform Memory Bank?

1. Set memory_revision_default_ttl to 90 days in the Memory Bank instance's TTL configuration.
2. Configure a granular TTL that sets only create_ttl to 90 days.
3. Schedule a Cloud Run job that lists memories daily and deletes the ones that are older than 90 days.
4. Set default_ttl to 90 days in the Memory Bank instance's TTL configuration.

Question 6, 3.1 Designing and building agentic workflows in code

An enterprise has 15 teams that build agents with Agents CLI, and they deploy from laptops and from CI. Governance requires that an SBOM and compliance check run before every deployment, and that coding agents learn the organization's agent-design rules from skills. Teams must keep typing the same agents-cli deploy command, and the platform team does not want to fork the tool. Which approach meets these requirements?

1. Build an Agents CLI extension that overrides deploy and ships the rule skills, add it at project scope, and commit its files.
2. Edit the installed google-agents-cli-deploy skill on each laptop so that it tells coding agents to run the compliance check before deploying.
3. Ask each developer to add the extension with --global on their laptop so that the override applies to all of their projects.
4. Fork the agents-cli repository, add the compliance check to the deploy command, and distribute the fork through an internal package index.

Question 7, 3.2 Integrating enterprise domain knowledge

A RAG Engine corpus created with the default RagManagedDb settings has grown from 3,000 to about 40,000 RAG files. Query latency has climbed steadily, and the product owner accepts a slight loss of recall in exchange for faster answers. The team does not want to take on the lifecycle and scaling of a separate vector database. What should you do?

1. Move the corpus to a Pinecone index configured as the RAG Engine vector database, so that a dedicated vector database scales with the larger volume.
2. Keep the default KNN strategy, and reduce chunk_size from 1,024 to 256 tokens so that each comparison in the exact similarity search is cheaper.
3. Create a corpus with the ANN strategy, tree_depth 3 and leaf_count near 10 times the square root of the file count, and import with rebuild_ann_index set to true.
4. Create a corpus with the ANN strategy and the default tree_depth of 2 and leaf_count of 500, and start querying at once because the index is built on the first query.

Question 8, 3.2 Integrating enterprise domain knowledge

A bank is building a RAG Engine corpus in europe-west4 for internal credit policies. Its security standard requires the corpus data at rest to be encrypted with customer-managed keys held in Cloud KMS. The project already has the Cloud KMS API enabled, and a security engineer holds the Cloud KMS Admin role. Which two actions are required to create the CMEK-protected corpus? (Choose two.)

1. Create the key ring in the global location so that the same key can protect RAG corpora in europe-west4 and in any other region.
2. Grant the RAG Engine service agent (service-PROJECT_NUMBER@gcp-sa-vertex-rag.iam.gserviceaccount.com) permission to use the key.
3. Reference the full key version resource name ending in /cryptoKeyVersions/1 in the corpus EncryptionSpec so that the key version is pinned.
4. Create an asymmetric signing key with the HSM protection level in europe-west4 so that every use of the key is hardware-backed.
5. Create a symmetric encrypt/decrypt key in a europe-west4 key ring, and reference it in the EncryptionSpec without the version suffix.

Question 9, 3.2 Integrating enterprise domain knowledge

A fraud-scoring agent runs on Agent Runtime with Agent Identity and reads a sensitive BigQuery dataset. A contractor proposes exporting the agent's Google Cloud access token to an on-premises batch job that needs the same data, and suggests disabling whatever blocks the token from working outside the runtime. The security team wants to prevent stolen or replayed agent credentials. What should you recommend?

1. Keep the default Context-Aware Access policy that binds agent tokens to the runtime, and give the batch job its own least-privileged identity.
2. Set the environment variable that opts the agent out of Context-Aware Access so that the batch job can reuse the agent's exported tokens on premises.
3. Generate a long-lived key for the agent identity, store it in Secret Manager, and let the on-premises batch job read the key when it starts each run.
4. Grant the batch job's service account the Service Account Token Creator role on the agent identity so that it can impersonate the agent at run time.

Question 10, 3.3 Orchestrating and coordinating agentic workflows

An insurer's claims-triage agent, built with ADK, must check three independent sources for every claim: the policy database, historical weather data, and a fraud-scoring API. Each check is handled by its own LlmAgent with dedicated tools. Running the checks one after another makes triage too slow. A final summary must combine all three results, and the orchestration must be deterministic rather than decided by a model. What should you do?

1. Run the three agents in a ParallelAgent with a distinct output_key each, inside a SequentialAgent that then runs a synthesis agent.
2. Put the three check agents and the synthesis agent together in one ParallelAgent so that all four agents start at the same time and the summary is ready sooner.
3. Wrap the three check agents in a LoopAgent with max_iterations set to 3 so that each iteration runs one check.
4. Create a coordinator LlmAgent with the three check agents as sub-agents and instruct it to transfer to each sub-agent before writing the summary.

Question 11, 3.3 Orchestrating and coordinating agentic workflows

Your underwriting agent is exposed through the A2A protocol, and some tasks take several hours to complete. The calling client is a serverless orchestrator running on Cloud Run functions. It cannot keep an HTTP connection open for hours, and the platform team wants to avoid running an extra always-on component just to check task status. How should you deliver task results to the client?

1. Have the client call SendStreamingMessage and keep the Server-Sent Events stream open until the task reaches a terminal state, resubscribing with SubscribeToTask if it drops.
2. Deploy a small Compute Engine VM that calls GetTask every few seconds for each open task and forwards completed results to the orchestrator through Pub/Sub.
3. Advertise capabilities.pushNotifications in the Agent Card, have the client register an HTTPS webhook through TaskPushNotificationConfig, and call GetTask on each notification.
4. Increase the client's HTTP timeout to several hours and use a synchronous SendMessage call for every request so that results return on the same connection.

Question 12, 3.3 Orchestrating and coordinating agentic workflows

A client agent asked a remote design agent, over the A2A protocol, to generate a campaign banner. The task reached the completed state and returned an image artifact. The user now wants the same banner with a different background color, and the design agent should keep the context of the original request. How should the client agent send this refinement?

1. Send a new message with the same taskId as the completed task so that the design agent reopens it and continues with the existing artifact and history.
2. Send a new message in the same contextId, listing the original task in referenceTaskIds, so that the design agent creates a new refinement task.
3. Start a new contextId and describe the original banner and the new color again in full so that the design agent has every detail it needs.
4. Call CancelTask on the completed task to release it, and then resend the original message with the color change in the same contextId.

Question 13, 3.3 Orchestrating and coordinating agentic workflows

An orchestrator ADK agent on Agent Runtime must discover a currency-conversion subagent from Agent Registry at runtime instead of using a hardcoded URL. The code calls AgentRegistry.get_remote_a2a_agent with the subagent's resource name. After deployment, the registry lookup fails with permission errors. In a test where the lookup is bypassed, calls to the Google-hosted subagent are rejected as unauthenticated. Which two actions should you take? (Choose two.)

1. Grant the Agent Registry API Viewer role (roles/agentregistry.viewer) to the developer's user account that deployed the orchestrator.
2. Grant the Agent Registry API Viewer role (roles/agentregistry.viewer) to the orchestrator's agent identity on the project or registry.
3. Pass a header_provider to the AgentRegistry constructor so that Google authentication headers are added to the remote A2A calls.
4. Pass an httpx.AsyncClient configured with Google authentication to get_remote_a2a_agent.
5. Store the subagent's endpoint URL in an environment variable and construct the RemoteA2aAgent from it.