Free CCAR-F Context Management & Reliability Practice Questions

The Context Management & Reliability domain makes up 15% of the Anthropic CCAR-F exam. Practice free questions covering context window management, compaction, prompt caching, retrieval, guardrails, evals, and human-in-the-loop review — each with the correct answer and a detailed explanation.

CCAR-F Context Management & Reliability Sample Questions with Answers

Sample Question 1 — Context Management & Reliability

A support agent has accumulated 70,000 tokens of chat history and tool responses while investigating a billing dispute. The agent must retain the exact disputed amount, transaction date, case ID, customer commitments, and unresolved issues. Which architectural change BEST reduces context pressure without risking loss of these facts?

  1. A. Retain every tool response and remove only the customer's older chat messages.
  2. B. Maintain a structured case record, trim processed tool output, and place the current case summary prominently. (Correct answer)
  3. C. Replace the entire history with a short narrative summary generated after each agent turn.
  4. D. Move the oldest messages into the middle of the prompt and preserve recent tool output.

Correct answer: B

Explanation: Correct answer (B): A structured case record preserves exact values and unresolved items while allowing verbose evidence already processed to be removed from active context. The underlying principle is to separate durable, decision-relevant state from transient conversational detail and keep critical state prominent. In production, this reduces token use and lost-in-the-middle failures without turning exact amounts, dates, or identifiers into vague prose. Why the other options are wrong: - Option A: Original tool output preserves evidence, but removing customer messages can discard commitments while retaining large volumes of irrelevant data. - Option C: Narrative summaries reduce context, but they can omit or blur exact amounts, dates, identifiers, and unresolved issues. - Option D: Burying older facts in the middle increases retrieval risk and does not meaningfully control context growth.

Sample Question 2 — Context Management & Reliability

An extraction pipeline processes 400-page insurance files. An upstream agent identifies policy values, and a downstream agent validates them against rules and citations. Passing every source page and the upstream agent's full reasoning exceeds the context budget. Source attribution must be preserved, and the validator must recognize missing fields. Which design is MOST appropriate?

  1. A. Send only a prose summary of the identified policy values and discard the original source pages.
  2. B. Divide the source into equal chunks and let the validator infer relationships without upstream findings.
  3. C. Send the complete source and omit upstream reasoning so all validation uses original evidence directly.
  4. D. Pass a structured fact manifest with values, source locations, and unresolved fields, retaining source text for targeted retrieval. (Correct answer)

Correct answer: D

Explanation: Correct answer (D): A structured manifest makes each extracted value, citation, and unresolved field explicit while allowing the validator to retrieve only the relevant source passages. The underlying principle is that downstream reliability improves when critical state is represented structurally and raw evidence remains available by reference. In production, this controls context size, preserves provenance, and prevents missing information from being mistaken for a confirmed absence. Why the other options are wrong: - Option A: A prose summary is compact, but discarding the source prevents citation verification and can hide unresolved fields. - Option B: Equal-sized chunks can separate related facts and force the validator to reconstruct findings without a reliable manifest. - Option C: Sending all 400 pages still exceeds the stated context budget and makes key fields difficult to locate.

Sample Question 3 — Context Management & Reliability

A support agent searches for an order using a customer's email address and finds two open orders with the same total but different delivery addresses. The customer has not identified which order has the problem, and changing the wrong order could interrupt delivery. What should the agent do FIRST?

  1. A. Ask the customer for a non-sensitive detail that distinguishes the two orders. (Correct answer)
  2. B. Select the most recently created order and disclose the assumption afterward.
  3. C. Escalate immediately because every duplicate search result requires human intervention.
  4. D. Update both orders so the requested change is guaranteed to affect the intended one.

Correct answer: A

Explanation: Correct answer (A): The agent should resolve the ambiguity with a targeted clarification before taking an action that could affect the wrong order. The underlying principle is to distinguish resolvable uncertainty from conditions requiring escalation. In production, explicit clarification prevents arbitrary record selection while avoiding unnecessary human handoffs. Why the other options are wrong: - Option B: Recency is not sufficient evidence when two plausible records exist and an incorrect change has consequences. - Option C: Escalation is appropriate when progress is blocked or policy requires it, but this ambiguity can likely be resolved through clarification. - Option D: Updating both records creates an unauthorized change to at least one potentially unrelated order.

Sample Question 4 — Context Management & Reliability

A market-research coordinator queries three specialist agents. One specialist receives HTTP 403 from a licensed database and currently returns an empty findings array. The coordinator then reports that no relevant acquisitions occurred. Which change BEST improves reliability?

  1. A. Have the specialist retry indefinitely until the licensed database returns a successful response.
  2. B. Have the coordinator treat every empty result as a failure and discard all completed research.
  3. C. Return a typed access failure with the attempted query, partial findings, and available alternatives. (Correct answer)
  4. D. Replace the failed specialist's output with a general conclusion generated from model knowledge.

Correct answer: C

Explanation: Correct answer (C): A typed failure lets the coordinator distinguish inaccessible evidence from a successful search that found nothing, while preserving partial work and possible recovery paths. The underlying principle is that multi-agent error contracts must carry failure semantics rather than collapsing failures into normal outputs. In production, this supports informed retries, alternate-source selection, and honest disclosure of coverage gaps. Why the other options are wrong: - Option A: A 403 commonly indicates an access or permission issue, so indefinite retries can stall the workflow without resolving the cause. - Option B: Legitimate searches can return no findings and must remain distinguishable from failed searches. - Option D: Model knowledge cannot replace required source access or justify presenting an unsupported conclusion as complete.

Sample Question 5 — Context Management & Reliability

A returns subagent calls an order lookup tool with a legacy invoice number. The tool reports a validation error and documents a safe alternative lookup by customer ID. If that lookup also fails, the coordinator must decide whether to escalate and must not claim the order was absent. Which implementation is BEST?

  1. A. Return the first validation error immediately and prohibit subagents from attempting alternative lookups.
  2. B. Try the documented alternative, then report both attempts, typed failures, and any partial results. (Correct answer)
  3. C. Retry the original invoice lookup several times, then return an empty order object.
  4. D. Ask the model to reconstruct the order details from conversation history and continue processing.

Correct answer: B

Explanation: Correct answer (B): The subagent can safely attempt the documented alternative locally and, if recovery fails, provide the coordinator with the failures, attempted actions, and partial results. The underlying principle is to recover locally when a safe path exists while preserving structured error information for higher-level decisions. In production, this avoids unnecessary escalation but prevents failed access from being misrepresented as proof that an order does not exist. Why the other options are wrong: - Option A: Immediate propagation preserves the original error but ignores an explicitly safe local recovery path. - Option C: A validation error requires changing the request, and returning an empty object would conceal the failure. - Option D: Conversation history is not authoritative enough to replace a required system-of-record lookup.

Sample Question 6 — Context Management & Reliability

A Claude Code workflow is tracing an authorization defect across a monorepo. After hours of exploration, the context contains full files, search logs, test output, hypotheses, and architectural notes. The investigation must continue across sessions and survive compaction. Which design BEST manages context?

  1. A. Keep every explored file in the main conversation and postpone compaction until the fix is merged.
  2. B. Compact immediately and rely on the generated summary to retain every relevant technical detail.
  3. C. Store only the current leading hypothesis and rerun repository searches whenever more evidence is needed.
  4. D. Use focused subagents and checkpoint decisions, evidence, open issues, files, and test status in a state manifest. (Correct answer)

Correct answer: D

Explanation: Correct answer (D): Focused subagents isolate exploration, while an explicit state manifest preserves confirmed findings, evidence, unresolved questions, affected files, and test status before context is compacted. The underlying principle is to externalize durable investigation state instead of relying on an ever-growing conversation. In production, this supports continuation, crash recovery, and efficient context reduction without losing the basis for engineering decisions. Why the other options are wrong: - Option A: Keeping all source material increases context pressure and can make important findings harder to retrieve. - Option B: Compaction is useful, but relying on it before explicitly preserving critical state risks losing precise evidence and unresolved work. - Option C: Dropping competing explanations, evidence, and completed searches encourages repeated work and confirmation bias.

Sample Question 7 — Context Management & Reliability

A repository-analysis session crashes after completing dependency tracing but before any code changes. A saved state manifest lists analyzed files, confirmed findings, unresolved assumptions, pending tests, and the next investigation step. What is the BEST recovery approach?

  1. A. Resume from the manifest and selectively reopen source files needed to verify the next step. (Correct answer)
  2. B. Repeat the entire repository exploration because no artifact can be trusted after a crash.
  3. C. Implement the proposed fix directly because every item recorded in the manifest is confirmed.
  4. D. Discard the manifest and ask a new session to infer prior progress from the issue title.

Correct answer: A

Explanation: Correct answer (A): The manifest provides an explicit checkpoint, and targeted source rereading verifies the evidence needed for the next action without repeating completed exploration. The underlying principle is to resume from durable state while distinguishing confirmed findings from unresolved assumptions. In production, this reduces duplicated work and prevents undocumented session memory from driving code changes. Why the other options are wrong: - Option B: Starting over is unnecessary when a usable manifest records progress and supports targeted verification. - Option C: Unresolved assumptions and pending tests must not be treated as confirmed facts before implementation. - Option D: The issue title lacks the evidence and progress necessary for reliable continuation.

Sample Question 8 — Context Management & Reliability

An invoice extraction service reports 97% aggregate field accuracy. However, handwritten freight invoices are rare, and incorrect bank account numbers create significant financial risk. The team wants a human-review policy that detects segment-specific weaknesses and routes uncertain fields appropriately. Which approach BEST satisfies these requirements?

  1. A. Review a simple random sample of documents and route entire documents using the model's raw average confidence.
  2. B. Review only documents the model labels low-confidence and calculate one accuracy score for the complete dataset.
  3. C. Sample by document type and field, calibrate field confidence on labeled data, and route risky uncertain fields. (Correct answer)
  4. D. Review every handwritten invoice permanently and accept all other documents without measuring field-level performance.

Correct answer: C

Explanation: Correct answer (C): Stratified sampling reveals performance on rare document types and sensitive fields, while labeled examples can calibrate field-level confidence for review routing. The underlying principle is that aggregate accuracy and raw confidence do not adequately represent segment-specific risk. In production, calibrated field-level review focuses human effort where errors are both likely and costly. Why the other options are wrong: - Option A: Rare invoice types and high-risk fields may be underrepresented or hidden by random sampling and document-level confidence. - Option B: Reviewing only low-confidence cases cannot reveal confidently wrong outputs, and a single aggregate score masks weak segments. - Option D: This does not calibrate confidence, assess other segments, or support adaptive routing.

Keep Practicing

Take the 10-question CCAR-F quick-start test across all 5 domains, or return to the CCAR-F practice test hub for 250+ questions, exam details, and a study plan. Drill the other domains: Agentic Architecture & Orchestration, Claude Code Configuration & Workflows, Prompt Engineering & Structured Output, Tool Design & MCP Integration, and Context Management & Reliability.