CCA Domain 5: Context Management & Reliability Explained

Domain 5 of the CCA-F exam, Context Management & Reliability, is 15% of the score and tests whether you can keep an agent accurate as its context fills: what summarization destroys and how to preserve it, when to hand off to a human, how errors should travel between components, how to detect degradation, and how to report findings with sources intact. The theme is that context is a budget with a shape, not a bucket.

This post covers what the domain includes, why it is weighted at 15%, the concepts to know cold, how the exam phrases them, a worked scenario, and the traps. It is built on our 31-article Domain 5 study guide; each idea links to its source article.

What does CCA Domain 5 cover?

Six task areas, with an anchor article for each:

Task	What it tests	Start here
5.1 Progressive summarization	What compression loses; case-facts blocks; lost-in-the-middle; tool output filtering	”$127.50 Refund” Became “Customer Requested a Refund” — The Agent Processed $50
5.2 Escalation triggers	Reliable vs unreliable triggers; explicit human requests; policy gaps; disambiguation	Sentiment-Based Escalation: 40% Volume, 30% Needed Human. Replace It.
5.3 Structured error context	Errors between orchestrator and sub-agents; silent swallow; coverage annotation	Generic “Failed” → 18% Recovery. Structured Error → 71%.
5.4 Context degradation	Specificity loss over time; scratchpads; delegation; crash manifests; /compact	Minute 0: “src/auth/jwt.ts:12 → verifyJWT()”. Minute 45: “Typical JWT Validation Pattern.”
5.5 Monitoring and accuracy	Aggregate metrics hiding failures; stratified sampling; field-level confidence; calibration	96% Accuracy, 40% More Escalations: Why Aggregate Metrics Lie
5.6 Provenance and source mapping	Claim-to-source attribution; conflicting data; temporal metadata; format matching	Claim-Source Mapping: Every Fact Needs a Return Address

The API-level background (the Messages API is stateless; you send the history every time) is in the Foundations primer, Tokens: The Currency of Every API Call, and is restated for this domain in The API Does Not Remember Your Conversation.

Why is Domain 5 weighted at 15%?

It is the smallest domain, but not because it is unimportant. The blueprint puts the building decisions in Domains 1 to 4 and reserves Domain 5 for what happens after the system has been running for forty turns or forty minutes. Fewer distinct mechanisms live here, so fewer questions, but they are the ones that separate a demo from a production system. On a 60-question form that is about 9 questions (confirm the current weights on the official Anthropic / Pearson VUE page before you register). Because Domain 5 leans on ideas from Domains 1 and 2, it is efficient to study last.

The concepts you must know cold

Summarization keeps sentiment and destroys numbers. Across summarised support conversations, sentiment survives at over 90% while exact amounts, dates and IDs survive at roughly a quarter. The fix is a structured case_facts block that persists next to the summary and is never compressed. Long multi-issue conversations need an issue tracker for the same reason (28% of Issues Dropped Without a Tracker. 3% With One.).

Attention is U-shaped. Sources at the start and end of a long input are incorporated at over 90%; sources in the middle at about half. Bigger context windows do not flatten the curve. Section headers plus a leading summary do (Sources 1-2: 96%. Sources 5-6: 52%. Sources 9-10: 94%.). Filter tool output to the fields you need before it lands in context (40 Fields Returned, 5 Needed), and have sub-agents return structured facts, not reasoning chains (2,800 Tokens of Reasoning Chain → 280 Tokens of Structured Facts).

Escalate on observable conditions, not on emotion or confidence. Explicit human request, detected policy gap, no progress after two attempts, and needed policy exception are reliable. Negative sentiment and low model confidence are not. When a customer asks for a person, escalate immediately (Immediate Escalation → CSAT 3.8). When two customer records match, ask, do not guess (Auto-Select “Most Recent”: 27% Wrong Customer. Ask for Email: 2%.). When policy is silent, escalate rather than decide (Policy Silent on Competitor Matching → Agent Decides).

Errors between components need structure too. A sub-agent that reports “database error” to the orchestrator triggers five pointless retries against a decommissioned system (The Database Was Permanently Decommissioned). Silent swallow and premature termination compound (Two Anti-Patterns That Compound). Reports must annotate what was not covered so a timeout does not read as “no impact” (The Search Just Timed Out).

Specificity decays with session time, independent of complexity. File:line references fall from over 90% in the first quarter hour to under a quarter after 45 minutes. Countermeasures: write findings to a scratchpad (The Scratchpad), delegate data-heavy work to sub-agents so the main context never holds it (Let Someone Else Hold the Data), keep a manifest for crash recovery (A Manifest File So You Don’t Start Over), and save before /compact, then restore (Save, Compact, Restore).

Aggregate accuracy hides the failures that matter. 97% overall can contain a 72% category. Disaggregate before deploying, sample by stratum (Stratified Sampling), route uncertain fields rather than whole documents to review (Field-Level Confidence), and calibrate what a confidence number actually means (The Model Says 0.9 — What Does That Actually Mean?).

Preserve conflicts; never average them. 35% from one source and 42% from another is not 38.5%. Present both with attribution and dates (Conflicting Data: Present Both, Fabricate Neither, Not Every Difference Is a Contradiction). Structure reports by evidence strength and match format to content type.

How does the exam test Domain 5?

Long-running scenarios with a symptom that emerged over time. Common shapes:

• “After a context summarization, the agent processed the wrong refund amount.” (Numbers were compressed away; add a persistent case-facts block.)
• “Escalations are 45% of volume and most were resolvable.” (Sentiment trigger; replace with observable conditions.)
• “The research report says no studies exist; 47 were found after the outage.” (Access failure reported as empty; annotate coverage, return structured errors.)
• “By minute 45 the exploration agent gives generic advice and contradicts earlier findings.” (Degradation; scratchpad and delegation.)
• “The extraction system passes at 96% but finance is fixing invoices every day.” (Aggregate masking; per-field and per-category metrics.)
• “Two sub-agents disagree on market share. What should the coordinator output?” (Both, attributed, with dates.)

Worked scenario: a 60-minute codebase audit that goes vague

Setup. An agent audits a 500-file repository for authentication issues. For the first fifteen minutes its findings cite file:line and function names. By minute 45 it says things like “the usual JWT validation pattern” and proposes refactoring a module it earlier described as well-structured. The team wants specific, consistent findings for the whole run.

Options.

A. Increase the model’s context window and rerun. B. Add “always cite file and line” to the system prompt. C. Have the agent write each confirmed finding to a scratchpad file as it goes, delegate per-directory reading to sub-agents that return structured findings, and rebuild from the scratchpad after any /compact. D. Split the audit into 60 one-minute sessions.

Reasoning. A does not change the attention curve or the accumulation of stale reasoning. B asks for specificity the context can no longer supply. D loses continuity entirely. C removes bulk data from the main context, persists specifics outside the window, and survives compaction. C is the least bad option, and it maps directly to the degradation article and its three follow-ups.

Common traps in Domain 5 questions

• Believing more context solves attention. Lost-in-the-middle persists at any window size.
• Treating summarization as neutral. It is lossy in a predictable direction; protect IDs, amounts and dates explicitly.
• Escalating on sentiment or confidence. Both are poorly correlated with actual need (High Confidence (0.9+): 12% Errors. Low Confidence (<0.5): 68% Correct.).
• Reporting “no results” for a failed lookup. Same trap as Domain 2, on the orchestrator side (Actually, 47 Papers Were Found After the Outage Ended).
• Random sampling for monitoring. Random samples mirror volume, so rare high-impact categories are barely checked.
• Resolving conflicts by recency or authority. The most recent write may be a bulk refresh; present both.
• Choosing prose for numbers or a table for analysis. Format follows content type (Tables for Numbers, Prose for Analysis).

Next step

Read the six anchor articles, then run the Domain 5 practice questions. Because this domain reuses ideas from Domains 1 and 2, do it after those two. When you are steady, sit the free CCA-F mock exam, which samples by official weight so Domain 5 appears in proportion. The complete list is on the Domain 5 study guide page, and the study-order companion is CCA Study Tips for Domain 5.