Workflow or agent
The central decision. A workflow executes a predetermined sequence. An agent lets the model decide the next step. Agents cost more, take longer, and fail in more ways, so the bar for choosing one is that the path genuinely cannot be known in advance.
The distractors in these questions are properties that sound decisive and are not. High volume does not favor either, because both scale. Calling several external APIs does not favor either, because both use tools. Variety in the input sources affects parsing, not orchestration.
From the free practice form
A developer is deciding between a fixed workflow and an autonomous agent for a document-processing pipeline. Which TWO characteristics of the task would most strongly favor implementing it as a workflow?
Context isolation is the pattern to know
Long-running agents degrade, and the exam has a preferred explanation for why. An orchestrator that accumulates raw source material fills its own window, and quality falls late in the run. The structural fix is to push the heavy material into subagents and have them return distilled findings, so the orchestrator holds conclusions rather than corpora.
From the free practice form
Harborline Logistics built a research agent that reviews shipping regulations across 14 jurisdictions in a single run. The orchestrator's context fills with lengthy source documents, and answer quality drops noticeably after the eighth jurisdiction. The team wants better late-run quality without moving to a model with a larger context window. Which change best addresses the problem?
Memory and state
Agent construction questions extend into state that outlives a single session, and the judgment is scope. State that belongs to one turn, one session, or one user across sessions are three different decisions, and putting long-lived state in the conversation is how an agent that worked on day one fills its window by day four.
How this domain interacts with the rest
Agents rarely appear alone. An agent question is often also a cost question, a tool-boundary question, or a security question about what the agent is permitted to do without a human. When two domains are plausibly in play, decide which constraint the stem actually binds on, and answer that one.