Total questions: 51
orientation-runtime-method
00-q1
Which statement best defines a production agent in this course?
Correct: b
The model is one component. The application owns validation, authorization, persistence, effects, observation, and recovery.
00-q2
Which requirements directly imply a durable workflow?
Correct: a, c
Long waits and late external signals require persisted execution state. Structured output is important but does not by itself require a durable workflow.
00-q3
What should an engineer do first when given ‘build an onboarding agent’?
Correct: c
Architecture follows the user outcome and consequence of failure, not a preferred framework.
should-this-be-an-agent
01-q1
Which use case most strongly justifies a bounded agent rather than a fixed workflow?
Correct: b
Dynamic action selection under uncertainty is a good reason for bounded autonomy. Deterministic rules should remain deterministic code.
01-q2
Which factors should reduce the amount of autonomy granted?
Correct: a, b, c
High consequence, irreversible effects, and clear deterministic logic all favor stronger workflow or human control.
01-q3
What is the best initial Atlas design?
Correct: b
Start with the minimum autonomy that produces measurable benefit and expand only after evaluation.
agent-loop-and-control-boundaries
02-q1
After a model returns a proposed tool call, what should happen next?
Correct: b
Tool availability in a prompt is not authorization. The application controls execution.
02-q2
Which are valid bounded-loop stop conditions?
Correct: a, b, c
A production loop has explicit success, budget, and no-progress termination or escalation.
02-q3
Why should a model return a reason code in addition to a free-form rationale?
Correct: b
Controlled categories are measurable and enforceable. They do not make model reasoning inherently truthful.
tool-contracts-and-side-effects
03-q1
A payment API times out after the charge may have succeeded. What is the safest next step?
Correct: c
This is an uncertain effect in the acknowledgment gap. Reconciliation prevents duplicate side effects.
03-q2
What belongs in a production tool contract?
Correct: a, b, c
A tool is an operational contract, not merely a prompt description.
03-q3
Which statement about ‘exactly once’ is most accurate?
Correct: b
Durable runtimes coordinate observations, but external activities may execute more than once.
context-engineering-and-grounding
04-q1
Where should ACL filtering occur for enterprise retrieval?
Correct: b
The model should never receive evidence the current principal is not authorized to access.
04-q2
Which metadata supports claim-level grounding?
Correct: a, b, c
Grounding needs traceable, versioned evidence, not a generic architecture label.
04-q3
A retrieved webpage instructs the agent to ignore policy. How should the system treat it?
Correct: b
Retrieved content may inform factual work but cannot become trusted system instruction.
state-sessions-and-memory
05-q1
Which store should preserve a large uploaded source document?
Correct: b
Large evidence belongs in artifact storage; run state should hold stable references and status.
05-q2
Which properties should a long-term memory write include?
Correct: a, b, c
Memory is governed data. Model extraction does not make a fact trustworthy.
05-q3
How is durable workflow state different from long-term memory?
Correct: b
Durability and memory solve different problems and have different correctness/lifecycle requirements.
orchestration-patterns
06-q1
When is parallel fan-out appropriate?
Correct: a
Independent safe work can run concurrently; dependencies and join requirements must remain explicit.
06-q2
What should a join policy define?
Correct: a, b, c
Production joins account for partial failure rather than assuming universal success.
06-q3
What most strongly justifies a separate specialist agent?
Correct: b
Separate agents should reflect a real system boundary, not role-play.
durable-long-running-agents
07-q1
Why should a model call be recorded as an activity result in a durable workflow?
Correct: b
Replaying control flow should not silently re-invoke a nondeterministic model.
07-q2
Which belong outside replayable workflow code?
Correct: a, b, c
External and nondeterministic work belongs in activities or runtime-provided recorded APIs.
07-q3
A workflow waits two days for approval. What resource should it consume while waiting?
Correct: c
Durable waits persist state and resume on an external event without holding compute.
human-in-the-loop-and-authority
08-q1
What must an approval be bound to?
Correct: b
A later edit or changed action must not inherit approval intended for an older version.
08-q2
Which are expected HITL outcomes?
Correct: a, b, c
Human review needs explicit paths, timeout/expiry, escalation, and durable state.
08-q3
Why is a UI approve button alone insufficient?
Correct: b
HITL is an end-to-end authority and state transition, not a front-end decoration.
multi-agent-mcp-a2a
09-q1
What is the safest interpretation of an MCP server advertising a tool?
Correct: b
Capability discovery does not grant authority.
09-q2
Which conditions may justify an A2A or independent-agent boundary?
Correct: a, b, c
Protocol boundaries add cost and failure modes and should map to real independence.
09-q3
How does a handoff differ from using an agent as a tool?
Correct: b
Responsibility, state, and user interaction differ between delegation styles.
security-identity-isolation-governance
10-q1
What is the strongest defense against a model generating another tenant’s record ID?
Correct: b
Security must hold even when model output is wrong or malicious.
10-q2
Which controls belong in a sandbox for generated code?
Correct: a, b, c
Isolation and least privilege constrain code even when generated content is malicious.
10-q3
Why is tool-output normalization important?
Correct: b
Normalization reduces instruction confusion but does not eliminate the need to validate or distrust external data.
agent-evaluation-and-testing
11-q1
Why can a correct final answer still be a failed agent run?
Correct: b
Production correctness includes process, policy, safety, and operations, not only output text.
11-q2
Which datasets should an agent evaluation program maintain?
Correct: a, b, c
A single tuned golden set creates overfitting and misses attacks and production failures.
11-q3
What is the preferred grader for a schema invariant?
Correct: b
Use the least ambiguous grader available; reserve model judges for genuinely semantic qualities.
observability-slos-and-incidents
12-q1
Which identifier should connect a multi-day Atlas execution across workers and callbacks?
Correct: b
Long-running causal correlation must survive process and service boundaries.
12-q2
Which conditions can be part of a valid-completion SLI?
Correct: a, b, c
HTTP success is insufficient when business, evidence, and authority invariants define success.
12-q3
Where should raw sensitive prompts usually be stored for debugging?
Correct: b
Prompts and outputs can contain customer data and require controlled access, retention, and redaction.
latency-cost-scale
13-q1
What is the best first latency optimization?
Correct: b
Reduce the work before choosing lower-capability paths or weakening controls.
13-q2
Which controls reduce retry-storm risk?
Correct: a, b, c
Retries must be bounded and coordinated to avoid multiplying load during dependency failure.
13-q3
Why might a cheaper model increase total cost?
Correct: b
Optimize total system economics using measured task quality, not model price alone.
deployment-enterprise-readiness
14-q1
What should identify the behavior of a production agent release?
Correct: b
Production behavior is created by the entire versioned system.
14-q2
What makes shadow execution safe?
Correct: a, b, c
Shadow runs compare behavior without allowing the candidate release to mutate production.
14-q3
A long-running workflow is active during rollback. What is the correct approach?
Correct: b
In-flight behavior must be deliberate and replay-compatible.
system-design-interview
15-q1
What should the first architecture diagram accomplish?
Correct: b
Start coarse and add detail only where constraints require it.
15-q2
Which clarifying questions usually change an agent architecture?
Correct: a, b, c
Authority, risk, and runtime shape determine core architecture.
15-q3
Which is the strongest tradeoff statement?
Correct: b
A strong tradeoff ties a choice to a requirement, names the cost, and states a trigger for change.
capstone-atlas
16-q1
Which Atlas action must remain outside autonomous model authority?
Correct: c
The capstone deliberately keeps the material business decision with an authorized human/system.
16-q2
Which demonstrations make the Atlas capstone more than a happy-path demo?
Correct: a, b, c
The capstone proves control, trust, durability, and human authority through failure.
16-q3
What should happen when the autonomous cost budget is exhausted before mandatory checks finish?
Correct: c
Budgets must not convert incomplete work into false certainty.