Developing custom agents

GCP-PAA · 66 questions

• A city 311 assistant needs to classify routine resident intents, like pothole reports versus noise complaints, at low latency and low cost. The team is weighing a self-hosted small language model against a larger proprietary LLM for this specific classification task. Which choice best fits the requirement?
• A county clerk's records agent handles sealed court filings and must satisfy a data-residency requirement that keeps this data within infrastructure the county directly controls. The team is deciding between a SaaS Gemini LLM call and a self-hosted open-source model. Which factor should drive the decision?
• A water utility's field-ops triage agent routes routine work-order requests through a smaller open-source model while reserving a large proprietary Gemini model for complex incident narratives. What principle does this architecture reflect?
• A housing authority builds an ADK-based eligibility agent that composes multiple typed tools, an income calculator and a document checker, rather than relying on a single monolithic prompt to handle the whole eligibility decision. What does this reflect about custom agent development with ADK?
• A transit agency's trip-planning agent, built with ADK, defines a typed schema for a fare-lookup function along with a callback that validates the tool's output before it's returned to the rider. What does this design correctly reflect about ADK tool integration?
• A permitting chatbot team must choose between a proprietary Gemini model accessed via Model Garden and an open-source model, given a policy that no applicant data may leave the agency's VPC. Which choice satisfies the constraint?
• An emergency-management alert agent uses a small on-device model for rapid triage during a network outage and a larger cloud-hosted LLM for detailed after-action narrative summaries once connectivity returns. What principle explains why this design is appropriate?
• A school-district IT helpdesk agent needs to preserve multi-turn context across a single ticket conversation. Instead of re-sending the full conversation history with every turn, the team configures managed sessions. What are managed sessions actually providing here?
• A public-health records assistant needs to recall a resident's prior inquiries across separate visits without improperly persisting protected health information. What memory configuration approach fits this requirement?
• A 311 assistant team is debating whether to store active conversation and session state in Firestore or in Cloud SQL. Given that this data is document-like, evolves per conversation, and doesn't need complex relational joins, which choice fits better?
• A library program assistant uses Memorystore for Redis as a fast short-term session cache while relying on Firestore for durable long-term user memory. What does this combination reflect about the two services?
• A parks-and-recreation reservation agent uses Agent Platform Memory Bank to remember a resident's favorite park and accessibility needs across unrelated future sessions, weeks apart. What does Memory Bank provide here that a session's transient state does not?
• A county IT team configures an ADK agent's compliance step to run in human mode via Agents CLI, requiring a staff member to approve a sensitive action before it executes, while routine steps elsewhere in the same workflow continue running in autonomous agent mode. What justifies this split?
• A housing authority adds a document-verification capability to its casework agent by installing an Agents CLI plugin, rather than rewriting the agent's core logic to add the capability directly. What does this reflect about how Agents CLI plugins work?
• A courts and clerk agent computes filing deadlines and must switch from autonomous agent mode to human mode via Agents CLI whenever a deadline calculation affects a party's legal rights, while continuing to run routine, non-binding date calculations autonomously. What threshold determines when the switch to human mode is required?
• A transit agency plans to host an agentic system on GKE integrated with sensitive SCADA controls for rail signaling, and infrastructure policy requires that no scenario training or inference data leave agency-owned infrastructure. Which consideration most directly narrows which language models the architect can even evaluate for this agent?
• A municipal open-data portal's agent sends simple metadata lookups, like the update frequency of the parking-sensor dataset, to a lightweight model, while complex cross-dataset analytical questions get routed to a larger Gemini model. What architectural pattern does this design represent?
• An emergency-management agency's incident-response agent must keep functioning during a regional internet or cloud-provider outage that could sever access to an external SaaS model API. The architect chooses a self-hosted open-source model running on agency-controlled infrastructure instead. What tradeoff does this decision primarily accept?
• A school district builds a tutoring agent by writing code with the Agent Development Kit (ADK), explicitly defining its system instruction and the exact set of tools it may call. A colleague asks why this wasn't configured instead in a low-code console like Gemini Enterprise's Agent Designer. What distinguishes the ADK approach for this use case?
• A city permitting agent walks an applicant through a multi-step application across several browser sessions. Weeks later, an abandoned application unexpectedly resumes with stale form data because the agent's session was left open indefinitely. What should the architect have configured to prevent this?
• A water utility's billing agent stores structured customer account and billing records in Cloud SQL, while keeping short-lived conversational turn context in Memorystore for Redis during an active session. What principle does this split reflect?
• A county procurement agent uses a large proprietary LLM to reason about high-stakes vendor-compliance determinations, but routes routine vendor FAQ questions to a smaller, cheaper model. What is the underlying selection principle?
• A permitting chatbot's RAG pipeline must index a large municipal zoning-code corpus organized into numbered sections and subsections with cross-references. Which embedding and chunking approach best fits this source material?
• A courts/clerk agent's retrieval step returns the top twenty case-law snippets by vector similarity, then a reranking model reorders those twenty before the top few are passed to the generation step. When does adding this reranking stage most clearly improve grounding compared to using the initial vector similarity ranking alone?
• A public-health records assistant's retrieval step is returning loosely-related patient records that lead the model to generate conclusions not actually supported by the source data. What configuration change most directly addresses this failure?
• A housing authority is weighing standing up and operating its own Vector Search index for case-file embeddings against using the managed Agent Retrieval service. Which factor should most influence that decision?
• A 311 resident-service assistant needs to ground its answers in a municipal knowledge base of ordinances, service schedules, and FAQ articles. The team chooses Agent Search, formerly Vertex AI Search, over building a custom Vector Search retrieval pipeline from scratch. When is this the sounder choice?
• A transit agency's trip-planning agent needs to read public schedule data, but the same Google Cloud project also stores internal HR records for agency employees. How should the architect scope the agent's Agent Identity to prevent the trip-planning agent from ever touching HR data?
• A water utility agent has two tools: one that creates work orders with write access, and one that looks up billing history and is strictly read-only, both governed by Agent Identity. What does scoping these two tools to different permission levels within the same agent accomplish?
• County IT builds a reusable address-validation capability and publishes it to Agent Registry so that agents built by the parks department, the permitting office, and the water utility can each discover and call it. What problem does this primarily solve?
• A school district's enrollment agent needs to query a Cloud SQL database of student records. The architect connects the agent to the database through the Google Cloud MCP Server rather than writing bespoke connection code inside the agent. What does this primarily provide?
• An emergency-management agency wants its incident-response agent to trigger alerts on a legacy siren and notification system that predates any modern API standard. No prebuilt MCP server or existing Google Cloud integration covers this system. When does building a custom MCP server for it make sense?
• A parks-and-recreation department's booking agent needs to check availability and reserve facilities through a third-party SaaS reservation platform that publishes a remote MCP server. Why would the department connect through that vendor-hosted remote MCP server rather than integrating directly against the platform's own API?
• A public library system has agents supporting catalog search, interlibrary loan requests, and patron account lookups, each built by a different branch's development team. The library's platform team catalogs a shared "check patron eligibility" skill in Skill Registry for all three agents to use. What does this cataloging step accomplish?
• A permitting agent's retrieval quality degrades sharply after a routine update: queries that previously returned highly relevant zoning-code passages now return near-random results, even though the documents themselves weren't changed. What most likely explains this?
• A housing authority's case-management system already has a prebuilt MCP Server that covers case creation, status updates, and document attachment. The architect chooses to connect the agent through that prebuilt MCP Server rather than building a custom API integration against the case-management system directly. When is this the sounder choice?
• A courts research agent needs to consult sealed case records as context when answering authorized staff questions, but must never modify those records. The architect scopes Agent Identity to grant read access to the sealed-records store and no write access at all. What does this asymmetric boundary accomplish?
• A 311 resident-service assistant retrieves ordinance passages by semantic similarity, but an outdated ordinance sometimes outranks a newer one that says essentially the same thing in fresher wording. The team adds a reranking step that weights recency alongside similarity. What does this reranking step accomplish?
• A transit agency is indexing its GTFS-derived service documentation for a rider-facing agent and must choose a chunk size for the retrieval corpus. What tradeoff does that chunk-size decision primarily involve?
• A county procurement office wants its vendor-sourcing agent to ground responses in vendor records stored in BigQuery, but the queries require custom filtering, joins, and business logic specific to the county's procurement rules. The team builds a custom integration layer between the agent and BigQuery rather than using a generic prebuilt connector. When does this custom layer make sense?
• A public-health department's clinical-coding agent needs similarity scoring tuned to structured diagnostic codes rather than generic document relevance. Which retrieval approach should the architecture team choose?
• A school district's engineering team needs a tool capability that another department already built and published. What should they do before writing any new code?
• A water utility exposes SCADA telemetry through an MCP server that must reject write commands from any agent. How should the team enforce this boundary?
• A housing authority's eligibility-policy retrieval pipeline must choose between a general-purpose embedding model and one tuned to housing-policy language. What consideration should drive the decision?
• A permitting department chains three agents so intake hands off to zoning-review, which hands off to approval, with each stage depending on the previous stage's output. Which orchestration pattern fits this design?
• An emergency-management team runs weather, traffic, and shelter-capacity agents whose outputs don't depend on each other, then aggregates the three results into one recommendation. Which orchestration pattern best fits this design?
• A 311 assistant must route a resident's request differently depending on whether it's a pothole report, a noise complaint, or a billing issue, with different follow-up steps for each path. Which orchestration pattern fits this requirement?
• A courts/clerk system's filing agent must hand off a case to a scheduling agent that belongs to a different team's separate Agent Runtime deployment. Which protocol is designed for this agent-to-agent handoff across separately owned deployments?
• A transit agency's trip-planning agent needs to call a shared fare-calculation tool hosted by another department. Which protocol fits this agent-to-tool access?
• A housing authority's triage agent needs to delegate to an eligibility-verification agent that it has no advance knowledge of at build time. How should it locate and delegate to that agent at runtime?
• A water utility's multiagent workflow needs one agent to be the only one that can issue work orders, while every other agent in the workflow may only read sensor data. What is the right way to enforce this difference?
• A school district's policy requires that a tutoring agent may invoke a grading agent only when the calling identity belongs to an authorized teacher-agent, not just any agent that knows the grading agent's capability. What does this scenario illustrate?
• A county IT team is deciding, for a specific integration, whether MCP or A2A is the correct protocol to use. What question should settle the choice?
• A permitting team runs a zoning check and a fire-safety check in parallel when they're independent, but runs them sequentially when the fire-safety check needs the zoning outcome as an input. What principle explains why the team switches patterns between cases?
• A public-health intake workflow needs to loop back to a clarification agent whenever the classification agent's confidence score is low, rather than proceeding with an uncertain result. Which orchestration pattern naturally supports this loop-back behavior?
• An emergency-management agent policy blocks a dispatch-recommendation agent from autonomously triggering an evacuation, requiring it instead to escalate the decision to a human-supervisor agent. What does this design choice reflect?
• A library/parks reservation agent hands off a booking to a separately hosted payment agent via A2A, and the requester's identity must carry through to the payment step. What must the handoff preserve for this to work correctly?
• A transit agency hosts a delay-predictor agent and a rider-assistant agent on Agent Runtime and coordinates them via A2A into one unified rider-facing chat experience. What does this design achieve?
• A housing authority's triage agent routes cases to specialized caseworker agents by category using a graph workflow, rather than having one generalist agent handle every case type itself. When does this specialization-plus-routing approach outperform a single generalist agent?
• A courts orchestrator agent selects the correct filing-type sub-agent by reading capability metadata published in Agent Registry, rather than the development team hardcoding a fixed lookup table of sub-agents. What benefit does this approach provide?
• A county procurement workflow enforces an audit trail by having each agent log its output before handing off to the next stage in a sequential chain. What does this design choice add to the workflow?
• A 311 assistant team is weighing a parallel-agent fan-out against a single agent performing sequential lookups for the same resident request. What tradeoff should drive their decision?
• A water utility requires an agent policy step that validates a calling agent's identity before granting cross-department sensor access over A2A. Where must that validation apply for the design to be effective?
• A school district's orchestrator agent discovers a newly published translation agent via Agent Registry and starts delegating to it without any redeployment of the orchestrator. What made this possible?
• An emergency-management graph workflow includes a retry/error branch that triggers when a downstream shelter-capacity agent times out. What does adding this explicit branch accomplish?
• A transit agency's rider-complaint workflow calls a fare API through a tool-access protocol for routine refund lookups, but when a complaint turns into a multi-department dispute it hands the case off to a specialized escalation agent using a different protocol built for agent-to-agent coordination. Which pairing correctly matches each protocol to its purpose in this workflow?