This domain is about 15% of the exam, roughly 9 of its 60 questions.

Every point below comes from Anthropic's official documentation; each section lists its sources.

Think first How do you keep exact amounts, dates and key findings intact as a conversation grows long?

Why it matters: Questions describe a long support or research session that loses details, and ask how to structure context so they survive.

Tool results use tokens out of proportion to their relevance, and models may omit findings from the middle of long inputs. Keep exact facts in a persistent block outside summarized history, and put key findings summaries first.

Term	Meaning
Case facts block	A persistent block of amounts, dates, order numbers and statuses in each prompt, kept outside summarized history.
Lost in the middle	Models handle the start and end of long inputs reliably but may omit findings from the middle.
Context layer	A separate store of structured issue data, such as order IDs and statuses, for multi-issue sessions.
Context editing	Removing old tool_result blocks from history once they are used, without restarting the conversation.

Put transactional facts in a persistent case facts block. Amounts, dates, order numbers and statuses go in a block included in every prompt, outside the summarized history. In multi-issue sessions, keep each issue's order ID, amount and status in a separate context layer.

Exam trap Do not rely on summarized history to carry exact amounts and dates. Keep them in a case facts block outside that history.

Tool results use tokens out of proportion to their relevance. An order lookup may return 40+ fields when only 5 are relevant. In long runs, tool definitions and piled-up tool results can exhaust the context window. Context editing removes old tool results once used, and programmatic tool calling keeps intermediate results out of history.

Put a key findings summary at the start of long inputs. Models handle the start and the end of a long input well, yet they may drop findings placed in the middle. Lead with the key findings, then organize the detailed results under explicit section headers.

Exam trap A key finding buried in the middle of a long aggregated input may be skipped. Place a summary of key findings at the beginning.

Context retention matters for conversation continuity. That includes a user who navigates away and later returns. In the support chatbot example, each user message is appended to a message list, and that whole list goes into the API call.

Have subagents return findings with their metadata. In the research exercise, each finding carries a claim, an evidence excerpt, a source URL or document name, and a publication date. The parent may summarize a subagent's report, so ask explicitly when you need it kept verbatim.

• Keep amounts, dates, order numbers and statuses in a case facts block in every prompt.
• Tool results can use tokens far out of proportion to their relevance, such as 40+ fields when 5 matter.
• Put key findings first and use explicit section headers in long inputs.
• Retaining context when a user leaves and comes back matters for continuity.

Sources: Claude Certified Architect – Foundations Exam Guide Version 1.0 · Effective July 2026 · Exam code: CCAR-F · This guide i · Manage tool context - Claude Platform Docs · Customer support agent - Claude Platform Docs · Subagents in the SDK - Claude Code Docs

Think first How should a support agent decide between resolving a case itself, asking a clarifying question, and handing it to a human?

Why it matters: Questions describe a support agent's escalation behavior and ask which trigger, prompt change or clarification fixes it.

Escalate on clear triggers: the customer asks for a human, policy has a gap, or the agent cannot progress. Sentiment and self-reported confidence do not measure complexity. When facts are ambiguous, ask rather than guess, and teach the boundaries with explicit criteria and few-shot examples.

Term	Meaning
Escalation trigger	A condition that sends a case to a human: a request for one, a policy gap, or no meaningful progress.
Policy gap	A request the policy is silent or ambiguous about, such as competitor price matching.
First-contact resolution	The support scenario's goal: 80% or more, while still knowing when to escalate.
Structured handoff	A summary for the human agent with customer details, root cause and recommended action.

Escalate for a human request, a policy gap or no progress. Policy exceptions and gaps are triggers, not just complex cases. Everything else the agent should try to resolve itself. In the guide's support scenario, the agent handles returns, billing disputes and account issues.

Common support situations and the right response

Situation	Response	Reason
Customer explicitly asks for a human	Escalate immediately	Honor the request without investigating first
Frustrated customer, issue within the agent's ability	Acknowledge, then offer to resolve	Escalate only if the customer repeats the request
Competitor price match, policy covers own-site only	Escalate	Policy is silent on this request
Lookup returns several matching customers	Ask for more identifiers	Do not pick one by heuristic

An explicit request for a human is honored at once. Do not investigate first when the customer clearly asks for a person. Frustration alone is different: if the issue is within the agent's ability, acknowledge the frustration and offer a resolution. Escalate if the customer then repeats the preference.

Exam trap Trying to solve the problem before honoring an explicit request for a human is wrong. The request itself is the trigger.

Sentiment and self-reported confidence do not measure complexity. Sentiment does not correlate with case complexity. An LLM's self-reported confidence is poorly calibrated, and the agent is already confidently wrong on hard cases.

Exam trap A confidence threshold sounds precise but fails, because the agent can be confidently wrong exactly where escalation matters most.

Add explicit escalation criteria and few-shot examples to the prompt. In the guide's sample, an agent at 55% first-contact resolution escalated simple damage replacements but handled policy exceptions alone. The root cause is unclear decision boundaries. Explicit criteria with examples are the proportionate first fix; a trained classifier needs labeled data and ML infrastructure.

Exam trap A separate classifier trained on historical tickets is over-engineered as a first step. Prompt optimization has not been tried yet.

Examples teach judgment that generalizes to new cases. Few-shot examples help the model handle novel patterns instead of matching only the listed cases. They also show how to treat ambiguous requests.

Ask for another identifier when a lookup returns several matches. Instruct the agent to request more identifiers rather than pick a match by heuristic.

Hooks can block policy-violating actions and route them to a human. For example, an interception hook can block refunds over $500 and redirect them to escalation. Human agents may not see the transcript, so the handoff summary gives customer ID, root cause, refund amount and recommended action.

• Escalate for an explicit human request, a policy gap or exception, or no meaningful progress.
• Honor an explicit request for a human immediately; for frustration alone, offer to resolve first.
• Sentiment and self-reported confidence are poor proxies for case complexity.
• With several matching records, ask for more identifiers instead of choosing by heuristic.
• Fix poor escalation calibration first with explicit criteria and few-shot examples in the prompt.

Sources: Claude Certified Architect – Foundations Exam Guide Version 1.0 · Effective July 2026 · Exam code: CCAR-F · This guide i

Think first When a subagent fails partway through a task, what should it tell the coordinator so the whole system recovers well?

Why it matters: Questions describe a failing subagent or coordinator and ask which error handling design lets the system recover.

A coordinator can only recover from errors it understands. Subagents fix transient failures locally and pass up only what they cannot resolve, with structured context and partial results. The coordinator then decides what to retry, reroute or report as a coverage gap.

Term	Meaning
Structured error context	Failure type, attempted query, partial results and alternative approaches, returned together with an error.
Valid empty result	A query that ran successfully but found no matches; it is not an error.
Coverage annotation	A note in synthesis output on which findings are well supported and which areas have gaps.
Transient error	A temporary failure, such as a timeout or service unavailability, unlike validation or permission errors.

Return the failure type, the attempt, partial results and alternatives. This structured context lets the coordinator make an informed recovery decision. Error metadata can also carry an errorCategory of transient, validation or permission, an isRetryable boolean, and a human-readable description. Marking errors retryable or not prevents wasted retries.

Exam trap A generic status such as "search unavailable" or "Operation failed" is not enough. It hides the context the coordinator needs to choose a recovery path.

Neither hide errors nor stop the whole workflow over one failure. Both are anti-patterns. Instead, pass the error up with structured context and let the coordinator proceed with partial results. The guide's timeout exercise tests exactly this.

Subagents handle transient failures themselves. Only errors a subagent cannot resolve go to the coordinator, together with what it attempted and any partial results. Claude Code itself makes up to 10 retries with exponential backoff on transient failures before it shows an error.

Transient, validation, business and permission errors need different handling. Timeouts and service unavailability are transient. Invalid input is a validation error, and a policy violation is a business error. For business rule violations, return retriable: false with a customer-friendly explanation, so the agent can tell the user what happened. MCP tools report failures with the isError flag.

Show which findings are well supported and where gaps remain. When a source was unavailable, the synthesis should say which topic areas have gaps. In the guide's exercise, a simulated subagent timeout still gives the coordinator structured context. It proceeds with partial results and marks the gaps. Conflicting credible sources keep both values with attribution.

In hub-and-spoke designs, the coordinator owns error handling. All subagent communication goes through the coordinator, which gives observability, consistent error handling and controlled information flow. It can evaluate synthesis for gaps and send targeted queries back to subagents until coverage is sufficient.

A run stopped at its turn limit comes back marked partial. maxTurns is the turn limit set for a subagent. When a subagent hits that limit, Claude Code flags its output as partial, which tells Claude the work is unfinished. An API error that ends a subagent early, such as a rate limit, is different: it is never delivered as the subagent's result.

• Errors that reach the coordinator carry failure type, attempted query, partial results and alternatives.
• Suppressing errors as empty success and ending a whole workflow on one failure are both anti-patterns.
• Subagents retry transient failures locally and propagate only what they cannot resolve.
• Synthesis output marks well-supported findings and the gaps left by unavailable sources.

Sources: Claude Certified Architect – Foundations Exam Guide Version 1.0 · Effective July 2026 · Exam code: CCAR-F · This guide i · Error reference - Claude Code Docs · Subagents in the SDK - Claude Code Docs

Think first How do you keep an agent accurate while it explores a large codebase over a long session?

Why it matters: Questions describe a long exploration that starts to drift or crashes, and ask which context technique fixes it.

Long sessions fill the context with verbose discovery output until answers drift. Keep the main agent's context lean: delegate exploration to subagents, write findings to files, summarize between phases, and persist state so work survives a crash.

Term	Meaning
Context degradation	Drift in long sessions: inconsistent answers and vague typical patterns instead of the specific classes found earlier.
Scratchpad file	A file where an agent records key findings and rereads them for later questions.
Manifest	A structured export of agent state that the coordinator loads on resume after a crash.
Subagent	A helper agent that works in its own context and returns only its final message to the parent.

Drift shows up as vague, inconsistent answers. In a long session, the model starts citing "typical patterns" instead of the specific classes it found earlier. Its answers also stop agreeing with each other. Scratchpad files of key findings help counteract this drift.

Let subagents absorb verbose output. A subagent's tool calls and results stay in its own context; only its final message returns. Spawn one for a specific question, such as finding all test files or tracing refund flow dependencies. The main agent keeps the high-level coordination.

Exam trap A subagent does not see the parent's conversation unless it is a fork, which inherits the full context. Otherwise, put the file paths, errors and decisions it needs into its prompt.

Write key findings to a file and reread them. An agent that records its discoveries in a scratchpad file can look them up for later questions instead of trusting fading context.

Summarize one phase before spawning the next. When an exploration phase ends, condense its key findings. Inject that summary into the initial context of the next phase's subagents, so they start from what is already known. A fresh start with a structured summary beats resuming with stale tool results.

Each agent exports its state to a known location. On resume, the coordinator loads the manifest and injects the saved state into agent prompts. Passing captured results into a fresh session is often more robust than relying on session resume.

Exam trap Resuming the old session is not the robust default after a crash. Load the structured state exports and build fresh prompts from them.

Run /compact when discovery output fills the context. During extended exploration, /compact reduces context usage once verbose output piles up. Compaction keeps the active context small, so pair it with memory for anything that must survive the summary.

Exam trap Compaction alone does not preserve every finding. Keep must-keep facts in memory so they outlast the summary.

• Inconsistent answers and talk of "typical patterns" signal context degradation.
• Delegate verbose exploration to subagents; only their final message returns to the parent.
• Persist findings in scratchpad files and phase summaries, and inject them into the next phase.
• For crash recovery, export state to manifests that the coordinator loads on resume.

Sources: Claude Certified Architect – Foundations Exam Guide Version 1.0 · Effective July 2026 · Exam code: CCAR-F · This guide i · Create custom subagents - Claude Code Docs · Subagents in the SDK - Claude Code Docs · Run agents in parallel - Claude Code Docs · Work with sessions - Claude Code Docs · Memory tool - Claude Platform Docs

Think first When is it safe to let high-confidence extractions skip human review?

Why it matters: Questions describe an extraction pipeline with limited reviewers and ask how to measure accuracy and route review.

A single overall accuracy number can hide weak document types or fields. Measure accuracy per segment, calibrate the model's field-level confidence against labeled data, and send only uncertain or conflicting cases to people.

Term	Meaning
Aggregate accuracy	One overall accuracy figure for all documents, which can mask weak document types or fields.
Field-level confidence	A confidence score the model outputs for each extracted field, used to route review.
Calibration	Setting review thresholds by checking confidence scores against a labeled validation set.
Stratified random sampling	Sampling high-confidence extractions to measure their error rate and spot new error patterns.

A 97% overall score can hide weak segments. One document type or field may perform poorly while the average still looks high. Before you reduce human review, break accuracy down by document type and by field. Automate only when every segment performs consistently.

Exam trap A high overall accuracy is not enough to automate extractions. Check each document type and field first.

Have the model score its confidence for each field. Field-level scores let you route review attention to the uncertain fields. Self-reported confidence is poorly calibrated on its own, though, so an agent can be confidently wrong on hard cases.

Exam trap Raw self-reported confidence is not a safe threshold. Calibrate it against a labeled validation set first.

Set thresholds against a labeled validation set. Compare the model's confidence scores with fields whose correct values you already know. Then pick the score below which a field goes to review, so attention lands where errors are likely.

Send low-confidence and conflicting cases to people. Route an extraction to a reviewer when the model's confidence is low, or when the source documents are ambiguous or contradictory. This spends limited reviewer capacity on the cases that most need it.

Keep sampling the extractions you automate. Use stratified random sampling of high-confidence extractions. It gives you an ongoing measure of their error rate and detects novel error patterns as they appear.

Exam trap Automating high-confidence extractions does not end measurement. Sampling them keeps their error rate visible.

Name what to report instead of asking for high confidence. An instruction like "only report high-confidence findings" improves precision less than specific categorical criteria. Say which issues to report and which to skip. Categories with many false positives also erode trust in the accurate ones.

• Validate accuracy per document type and field before cutting human review.
• Calibrate field-level confidence thresholds with a labeled validation set.
• Route low-confidence or contradictory-source extractions to human reviewers first.
• Keep stratified random sampling running on high-confidence output to track errors.

Sources: Claude Certified Architect – Foundations Exam Guide Version 1.0 · Effective July 2026 · Exam code: CCAR-F · This guide i

Think first How does a multi-agent research system keep track of where each claim came from, and what should it do when sources disagree?

Why it matters: Questions describe a research pipeline that lost citations or mishandled conflicting data and ask which design fixes it.

Every claim should travel with its source from the first agent to the final report. Summaries that drop that link lose attribution. When sources disagree, the system keeps both values, labeled, and lets the coordinator decide how to reconcile them before synthesis.

Term	Meaning
Claim-source mapping	A structured link from each claim to its evidence, URL or document name, and date.
Provenance	Where a claim came from, kept with the claim as metadata through each step.
Contested finding	A finding that credible sources disagree on, reported apart from settled ones.
Coverage annotation	A note showing which findings are well supported and where sources were missing.

Summaries drop sources when claims lose their mappings. Each time findings are compressed, the link from claim to source can disappear. Numbers, percentages and dates can also blur into vague wording.

Exam trap A shorter summary is not a safer one. Compressing findings without their claim-source mappings is exactly how attribution gets lost.

Keep content and metadata in separate fields. Each finding carries a claim, an evidence excerpt, a source URL or document name, and a publication date. The synthesis agent must preserve that attribution when it merges findings.

What a structured finding carries

Field	Kind	What it protects
Claim and evidence excerpt	Content	What was found and what supports it
Source URL or document name	Metadata	Attribution through every later step
Publication or collection date	Metadata	Correct reading of changes over time

Require publication or collection dates in every output. Two figures from different years may both be right for their time. Without dates, that difference can be misread as a contradiction.

Report both values with their sources; never pick one at random. When two credible sources disagree, annotate the conflict. Document analysis should finish with both values marked, and the coordinator decides how to reconcile them before synthesis.

Exam trap Choosing the more recent or more likely value silently is the wrong move. Keep both values, each with its source, and flag the conflict.

Give well-established and contested findings their own sections. Keep each source's own characterization and its method context. Coverage notes can also show where gaps exist because sources were unavailable.

Tables for financial data, prose for news, lists for technical findings. Give each content type the form that suits it, instead of one uniform format for everything in the synthesis.

• Attribution is lost when findings are summarized without their claim-source mappings.
• Each finding should carry a claim, evidence excerpt, source URL or document name, and date.
• When credible sources conflict, keep both values with attribution instead of choosing one.
• Dates in structured outputs stop differences over time from looking like contradictions.

Sources: Claude Certified Architect – Foundations Exam Guide Version 1.0 · Effective July 2026 · Exam code: CCAR-F · This guide i