PREPARING FOR CERTIFICATION

SAMPLE QUESTIONS Choose the best answer for each question, then check the answer key that follows each one. • In the ETCLOVG taxonomy, which layer includes permission gates and approval steps? A. Execution. B. Context. C. Governance. D. Observability. ◦ Answer: C. ◦ Governance covers safety, security, and compliance, which includes permissions and approvals. • What are the three beats of the basic agent loop? A. Plan, build, ship. B. Think, act, observe. C. Read, write, delete. D. Train, test, deploy. ◦ Answer: B. ◦ The harness asks the model what to do, runs the tool, and shows the model the result. • A conversation has grown so long that the model misses facts from the middle of it. This is most closely described as which problem? A. Prompt caching. B. Context rot. C. Fan-out. D. Idempotency. ◦ Answer: B. ◦ Context rot is the loss of reliable attention as context grows. • Which action is idempotent? A. Appending a line to a log file. B. Sending an email. C. Setting a report to a fixed text when it already holds that text. D. Adding one to a counter. ◦ Answer: C. ◦ Doing it again leaves the system in the same state. • What does the principle of least privilege recommend for a tool that reads resumes? A. Allow it to read the whole disk in case it is needed. B. Allow it to read only the resumes folder. C. Allow it to read and write anywhere. D. Allow it to run any command. ◦ Answer: B. ◦ Grant only the access the job requires. • Which statement best distinguishes graceful degradation from graceful failure? A. They are identical. B. Degradation keeps operating with reduced function, while failure ends a single run cleanly with a clear reason. C. Degradation is for people, failure is for machines. D. Failure keeps the system running. ◦ Answer: B. ◦ The first reduces function while continuing, and the second is a clean ending. • A resume contains hidden text telling the model to ignore its instructions. This is an example of what? A. Tool poisoning. B. Prompt injection. C. Context compaction. D. Model routing. ◦ Answer: B. ◦ Orders hidden in data the agent reads are prompt injection. • A tool description contains a hidden order to secretly read every file. This is an example of what? A. Tool poisoning. B. A runaway loop. C. Backoff. D. A golden answer. ◦ Answer: A. ◦ Hidden orders inside a tool description are tool poisoning. • What is a trace? A. A single timed piece of work. B. All the spans belonging to one task. C. A list of passwords. D. A copy of the model. ◦ Answer: B. ◦ A span is one piece of work, and a trace gathers the spans of a task. • Which sequence of waits is exponential backoff? A. 1, 1, 1, 1 seconds. B. 1, 2, 3, 4 seconds. C. 1, 2, 4, 8 seconds. D. 8, 4, 2, 1 seconds. ◦ Answer: C. ◦ The wait doubles after each failure. • Which three case types should an evaluation set include? A. Fast, slow, and medium. B. Normal, edge, and adversarial. C. Cheap, strong, and hybrid. D. Read, write, and delete. ◦ Answer: B. ◦ Normal cases, awkward edge cases, and attacks. • What is the main benefit of giving a writer subagent only a small structured summary? A. Context isolation. B. Longer prompts. C. More tools. D. Faster typing. ◦ Answer: A. ◦ The subagent sees only what it needs, which keeps its context clean. • Which hook runs just before a tool and can refuse the call? A. Post-execution hook. B. Pre-execution hook. C. Logging hook. D. Shutdown hook. ◦ Answer: B. ◦ A pre-execution hook runs first and may change or refuse the request. • Which known weaknesses can affect an LLM used as a judge of other outputs? A. Position bias and verbosity bias. B. Memory leaks. C. Slow disks. D. None. ◦ Answer: A. ◦ Judges can favor answers by their position or length, so test them against human judgments. • The MCP is best described as what? A. A database. B. An open standard for connecting models to tools and data. C. A model training method. D. A file format for resumes. ◦ Answer: B. ◦ It is an open standard connecting models to tools and data. • What is the main risk of careless compaction? A. It makes the model faster. B. It can silently lose an important decision. C. It deletes the logs. D. It raises the temperature. ◦ Answer: B. ◦ A careless summary can drop facts, so important decisions must be preserved. • Why does an agent loop need a step limit? A. To make the model smarter. B. To guarantee the loop ends even if the model never says it is finished. C. To reduce the number of tools. D. To hide errors. ◦ Answer: B. ◦ The limit guarantees an ending and caps cost. • In RAG, what happens? A. The model is retrained on every question. B. Relevant documents are looked up and given to the model along with the question. C. The model deletes old documents. D. The harness writes the answer without the model. ◦ Answer: B. ◦ Retrieval finds relevant material and adds it to the context. • Which is the best reason to log a security event when an injection is detected? A. To slow the system. B. To create a record that supports investigation and regression tests. C. To confuse attackers. D. To save disk space. ◦ Answer: B. ◦ A record lets you investigate and add a regression test. • After fixing one failing evaluation item, what should you do next? A. Stop, because the item passes. B. Re-run the entire evaluation set. C. Delete the failing item. D. Lower the pass mark. ◦ Answer: B. ◦ A fix can break something else, so run the whole set.