This quiz is interactive and needs JavaScript. Here are the questions for reference:
1. Anthropic draws one architectural line between "workflows" and "agents." Which statement matches that line?
- A) Workflows use tools; agents do not
- B) Workflows orchestrate LLMs/tools through predefined code paths; agents let the LLM dynamically direct its own process and tool usage
- C) Workflows are stateless; agents are stateful
- D) Workflows call one model; agents call many models
2. A task splits cleanly into three independent subtasks whose nature you know in advance, and you run all three simultaneously then aggregate. Which pattern is this?
- A) Orchestrator-workers
- B) Prompt chaining
- C) Parallelization (sectioning)
- D) Routing
3. What is the single testable distinction between the orchestrator-workers workflow and the parallelization workflow?
- A) Orchestrator-workers runs sequentially; parallelization runs concurrently
- B) Orchestrator-workers determines its subtasks dynamically at runtime; parallelization uses pre-defined subtasks
- C) Orchestrator-workers uses one model; parallelization uses many
- D) Orchestrator-workers has no aggregation step
4. An incoming support message is classified into a category and sent to a specialized prompt built for that category. Which workflow is this, and what does it enable?
- A) Evaluator-optimizer; iterative refinement
- B) Prompt chaining; fixed decomposition
- C) Routing; separation of concerns / more specialized prompts
- D) Parallelization; voting
5. Per Anthropic, what is the "Augmented LLM" — the basic building block of every agentic system?
- A) An LLM fine-tuned on your domain data
- B) An LLM augmented with retrieval, tools, and memory that the model actively uses
- C) A router in front of several specialized models
- D) A generator paired with an evaluator in a loop
6. Which scenario is the textbook signal to reach for an autonomous agent rather than any fixed workflow?
- A) The task cleanly decomposes into fixed subtasks
- B) Inputs fall into a few distinct, accurately-classifiable categories
- C) It's an open-ended problem where you can't predict the number of steps or hardcode a fixed path
- D) You need multiple perspectives on the same task for higher confidence
7. How does Anthropic define context engineering relative to prompt engineering?
- A) They are synonyms
- B) Prompt engineering is curating tokens across turns; context engineering is writing a single instruction
- C) Context engineering is the broader discipline of curating/maintaining the optimal set of tokens across turns; prompt engineering is a subset (writing/organizing instructions)
- D) Context engineering only concerns the system prompt; prompt engineering concerns everything else
8. A system prompt hardcodes brittle, complex if/then logic for every edge case. Per the "altitude" framing, which failure mode is this and what is the fix?
- A) Too high / over-general; add more rules
- B) Too low / over-specific; prefer strong heuristics plus a few canonical examples
- C) Correct altitude; leave it
- D) Too high / over-general; make it vaguer
9. In MCP, which trio of primitives is exposed by servers (not clients)?
- A) Sampling, Elicitation, Logging
- B) Tools, Resources, Prompts
- C) Tools, Sampling, Roots
- D) Resources, Prompts, Elicitation
10. You are configuring MCP for a deployment. A local server runs as a subprocess on the same machine; a remote server is shared by many clients over the network. Which transport does each typically use, and where does SSE fit?
11. Context editing (clear_tool_uses_20250919) requires the beta header context-management-2025-06-27, whereas the memory tool (memory_20250818) is generally available and needs no beta header.
12. Structured note-taking and server-side compaction are the same technique: both summarize the in-context message history.
13. Name the two variations of the parallelization workflow and the primary benefit each targets.
14. State, in one sentence, the guiding principle that governs when to add agentic complexity (the "start simple" rule that appears across the workflow/agent notes).
15. Your agent processes a large corpus. Instead of pre-loading everything via embeddings, you have it keep file paths and stored queries as lightweight identifiers and load data on demand with tools. Which strategy is this, and what is its main trade-off versus pre-loading?
16. You want to systematically improve a tool suite. You generate dozens of realistic multi-tool prompt/response pairs, run each through a simple while-loop agentic loop, and collect metrics. You notice many redundant tool calls and several invalid-parameter errors. Per the evaluation-loop guidance, what fix does each metric point to?