Independent: not affiliated with Google Cloud. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its Google Cloud source.
5 questions / Explanations link Google Cloud sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Section 1 / Task 1.1 Configure agentic workflows and behavior with low-code tools
A parcel carrier's CX Agent Studio agent application hands the customer to a flow that collects a tracking number and a new delivery slot in a fixed order. The agent already holds the customer's account ID before the handoff, and the flow needs it to confirm that the tracking number belongs to that customer. Once the flow ends, the agent's rebooking tool needs the two values the flow collected. How should the values move between the agent and the flow?
Show answerHide answer
Answer: D
Why
CX Agent Studio agents treat flows as encapsulated black boxes: an input variable mapping sends the agent's variables to the flow as session parameters, and an output variable mapping sends the flow's session parameters back to the agent. The account ID goes in; the two collected values come back.
Why not the others
AA webhook detour through Firestore is not the handoff mechanism the page describes, and it adds a datastore and an extra call that the input and output mappings make unnecessary.
BThe page says to use CX Agent Studio agents as the steering agents that manage routing between agents and flows, and the conversation returns to the agent that started the flow when the flow reaches END_SESSION.
CThe page has agents treat flows as encapsulated black boxes, so everything the flow needs, including the account ID, reaches it through explicit input parameters rather than by the flow reading the agent's variables.
Section 2 / Task 2.1 Use coding agents effectively
A tech lead assigns an Antigravity agent a change to a payments module that will touch dozens of files. Before the agent writes to any of them, the lead wants to read the agent's approach for the whole change, mark it up with feedback, and have the agent revise it until the lead signs off. Which Antigravity feature supports this?
Show answerHide answer
Answer: A
Why
The implementation plan is the artifact the agent uses to architect changes in the codebase; unless the review policy is Always proceed, the agent requests review of the plan before changing code. The lead submits comments with Review, and the agent revises until the lead clicks Proceed.
Why not the others
BDefault mode pauses on each file write to show a diff to accept or reject, which reviews individual edits one at a time, not the approach for the whole change.
CA walkthrough is created when the agent completes task implementation; it summarizes changes that have already been made.
DA subagent in an isolated Git worktree does its editing on its own branch first, so the lead's review of that branch comes after the edits, not before.
Section 3 / Task 3.1 Design and build agentic workflows in code
Four teams run ADK agents that each bundle their own copy of a returns-policy skill, and the copies have drifted apart. The company wants a policy update to reach every agent from one place, and wants every earlier version of the skill kept unchanged so that auditors can look at it later. Which approach fits?
Show answerHide answer
Answer: B
Why
Skill Registry is a private repository for agent skills: agents discover and load a skill from it at run time, so one registry skill replaces the four copies. Each revision is an immutable snapshot of a skill version, which keeps earlier versions unchanged for audit.
Why not the others
AMemory Bank holds memories generated from conversations with a user; it is not a repository for skill packages that agents load.
CFour servers are four more copies to keep in sync; the requirement is one source with every earlier version preserved.
DA line in an instruction names a version but synchronizes nothing; the drift comes from maintaining four copies, and each copy can still be edited on its own.
Section 4 / Task 4.1 Evaluate agents in development and production
A team runs an offline evaluation of a travel-booking agent on 1,500 cases, scored with the Agent Multi-turn Task Success and Agent Multi-turn Tool Use Quality metrics, and 240 cases fail. The team has one sprint and wants to spend it on the few changes that remove the most failures. What should it do first?
Show answerHide answer
Answer: C
Why
Automatic Loss Analysis classifies the failures from rubric-based metrics into loss patterns, shown as Failure Clusters in the results or generated with generate_loss_clusters, with taxonomies for the two metrics named here. The triage workflow then uses the pattern category to pick a targeted fix and re-runs the evaluation.
Why not the others
AOnline monitors score production traffic over time; they do not group this offline run's failures by cause.
BThe triage workflow reads traces after the clusters, to find the exact turn or tool call inside a pattern; going through 240 traces one at a time finds individual defects but not which patterns account for the most failures.
DA larger model is a change made before anyone knows why the cases failed; the triage workflow finds the cause from the loss patterns first and then applies a targeted fix to the system instructions, tool definitions or few-shot examples.
A support agent on Agent Runtime calls a CRM MCP server through Agent Gateway, and some tool responses contain customers' payment card numbers. Security requires that card numbers be removed from a tool response before it reaches the model while the rest of the response still gets through, and that a response which also carries a prompt injection attempt be blocked outright. Which Model Armor configuration meets both requirements?
Show answerHide answer
Answer: D
Why
Tool responses reach the agent through the Agent-to-Anywhere (egress) gateway, and redaction needs advanced Sensitive Data Protection settings with a de-identify template and the enforcement type INSPECT_AND_BLOCK; with INSPECT_ONLY, Model Armor does not redact. A payload that violates several detectors is blocked whole rather than redacted.
Why not the others
AAn instruction is guidance the model may not follow; the card numbers would still reach the model.
BThe ingress gateway screens requests from clients to the agent, basic Sensitive Data Protection supports inspection only and no templates, and INSPECT_ONLY does not redact.
CWith INSPECT_ONLY, Model Armor logs findings but does not redact the payload or stop it.
The study bundle has two practice exams with 120 original questions, an answer key and an explanation for every answer, plus the 58-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
120 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.