CCAR-P Advanced Agentic Architecture Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Summarize and discard the oldest transcript turns once a threshold is reached, retaining only the most recent turns verbatim.

  Why it's wrong here

  Naive truncation of the oldest turns discards the very findings that later decisions depend on, causing the agent to contradict itself or repeat completed work. Summarization at the boundary helps only if the summary is curated and persisted; simply dropping turns loses causal context. This approach trades context overflow for amnesia, which is worse for a long research task.

• ✗

  Enable extended thinking on every turn so the model can reason through the full history internally without needing external state.

  Why it's wrong here

  Extended thinking improves the quality of reasoning within a single turn but does not expand the input context window or persist knowledge across turns. Once the transcript is compacted, earlier thinking blocks are gone. Relying on internal reasoning alone leaves the agent dependent on whatever survived truncation, so it cannot guarantee continuity of findings across the boundary.

• ✗

  Raise the max_tokens parameter on every request so the model can hold more of the transcript in a single response.

  Why it's wrong here

  max_tokens governs how many tokens the model may generate in a response, not how much input context it can accept. Increasing it does not expand the context window and may even increase cost and latency. The transcript will still overflow the input limit, so this does not address the compaction problem and could make responses unwieldy.

• ✓

  Maintain an external structured scratchpad of confirmed findings, open questions, and decisions, and inject a curated summary of it into each new context window.

  Why this is correct

  An external scratchpad decouples durable knowledge from the ephemeral transcript, so compaction can drop raw turns without losing conclusions. Injecting a curated summary re-establishes the essential state each cycle, and because the scratchpad is structured, the agent can update specific fields rather than rewriting prose. This keeps decisions consistent even when the underlying conversation is truncated.

• ✓

  Persist a durable task state object and rehydrate the agent from it at each context boundary, treating the transcript as a disposable execution log.

  Why this is correct

  Treating the transcript as disposable and the task state as durable inverts the usual assumption and makes compaction safe. The agent reloads goals, constraints, and accumulated findings from the state object, so no critical decision depends on a turn that may be evicted. This also makes the workflow resumable after crashes or restarts, which is valuable for multi-hour runs.