Claude conceptsClaude Certified Architect – FoundationsContext management
CCAR-F context management, escalation and human review
What CCAR-F asks about keeping long contexts reliable, persisting case facts, and which signals mean an agent should escalate to a person.
A single order lookup can return 40+ fields when only 5 are relevant, and those results pile up in context. Routing to humans by the agent's own confidence score fails, because that confidence is poorly calibrated.
01 What does CCAR-F test about context, escalation and human review?
You read a long support transcript in a question and feel the middle details slipping. The model has the same weakness: it reliably processes the beginning and end of long inputs but may omit findings from the middle.
Expect questions on structuring context, deciding when to escalate versus resolve autonomously, and human-in-the-loop patterns.
Put what matters at the start of long inputs, where the model reads it reliably.
02 How do you keep long contexts reliable?
Place a key findings summary at the beginning of aggregated inputs and organize the details under explicit section headers. Keep amounts, dates, order numbers and statuses in a persistent "case facts" block in each prompt, outside summarized history.
Trim what you no longer need: context editing removes old tool_result blocks without restarting the conversation. Subagents help too, since only their final message returns to the parent.
Keep transactional facts in a case facts block, not in summarized history.
03 When should an agent escalate to a person?
When to escalate to a person: the signals compared
Escalate
- The customer asks for a human
- Policy is silent or needs an exception
- The agent cannot make meaningful progress
Resolve or ask
- Frustrated customer, issue within reach
- Several matches: ask for another identifier
- Standard damage claim with photo evidence
Not a reliable signal
- Customer sentiment
- Self-reported confidence score
Honor an explicit request for a human immediately, without investigating first. For a frustrated customer whose issue the agent can handle, acknowledge the frustration and offer a resolution, escalating only if they repeat the request.
In the guide's sample, an agent at 55% first-contact resolution, against an 80% target, escalated simple cases and tried to handle policy exceptions itself. The fix is explicit escalation criteria with few-shot examples, before any classifier.
Unclear decision boundaries cause bad escalation; explicit criteria fix them.
04 Where does human review fit?
Automated review can stay advisory. In Claude Code Review, a verification step filters out false positives, and a review never approves or blocks a pull request, so the merge decision stays with your team.
Automated review informs people; the decision stays with them.
Summary
- Put key findings first and details under explicit headers.
- Keep amounts, dates, order numbers and statuses in a case facts block.
- Escalate on customer requests, policy gaps and lack of progress.
- Sentiment and self-reported confidence are poor escalation signals.
Sources and editorial status
Research checked Nov 3, 2026.