Anthropic CCDV-F Claude Certified Developer-Foundations Exam Practice Test

The supplied examination source identifies D as correct. The statement "produce client memos faster" is a business objective , not an implementable engineering requirement. The first engineering activity is therefore requirements decomposition: determine what functionality is required and what technical characteristics must support it.

Functional requirements might specify how analysts provide source information, what stages the agent performs, whether it researches, outlines, drafts, revises, or cites material, which systems it accesses, what output it returns, and where human approval occurs. Infrastructure and non-functional requirements then define latency, access control, confidentiality, context size, availability, auditability, model selection, integration interfaces, token cost, observability, and operational constraints.

Anthropic's developer lifecycle similarly separates defining measurable success criteria from choosing and implementing model capabilities. Its evaluation guidance emphasizes establishing specific, measurable success criteria before optimizing a model or prompt. Claude's agent configuration also separates implementation components such as the model, system prompt, tools, MCP servers, and Skills—choices that logically follow requirement definition.

A is useful discovery but prematurely assumes a prompt-driven drafting implementation. B is precedent-driven rather than requirement-driven. C selects technology before defining what the system must accomplish.

Relevant topics: Understanding Reqs, functional requirements, infrastructure requirements, non-functional requirements, acceptance criteria, and solution decomposition .

The supplied Claude Certified Developer Foundations material identifies A as the best answer. The decision should begin with the actual capability gap rather than a predetermined preference for built-in or custom tooling. If the existing capability can satisfy the requirement with a small supported extension or composition, using it avoids unnecessary implementation and maintenance. If the missing capability is fundamental, a custom tool becomes justified.

Anthropic's tool architecture explicitly supports both categories. Claude provides Anthropic-managed or predefined tools while also allowing developers to define custom tools for operations specific to their application. With a custom client tool, the application defines the contract and executes the operation before returning its result to Claude.

A is more complete than C because it evaluates both sides of the boundary : whether the existing mechanism can realistically meet the missing requirement and, if not, whether custom implementation is warranted. B prioritizes precedent over current requirements. D only determines whether the unmodified built-in tool is sufficient and does not address what to do when it is almost—but not entirely—sufficient.

Relevant Claude Developer topics: built-in tools, custom tools, capability-gap analysis, tool contracts, maintenance cost, extensibility, and agent construction decisions .

===============

The correct approach is to evaluate each customization independently against both its original requirement and the capabilities currently supplied by the Claude SDK. Custom code has an ongoing maintenance cost, but automatically removing it is unsafe because some modifications may encode genuine application requirements that the standard SDK behavior still does not address. Conversely, automatically retaining every customization can preserve obsolete complexity after the SDK has evolved.

This is particularly relevant because Anthropic's SDKs already provide capabilities that developers might otherwise implement themselves. Official Claude API documentation states that SDKs provide built-in request formatting, error handling, retries, timeouts, and other integration functionality. For example, transient connection, rate-limit, and server errors are retried automatically with exponential backoff.

Therefore, each custom history, retry, or error-management layer should be reviewed for its actual purpose. If an SDK primitive now provides equivalent behavior, removing duplicate code improves maintainability. If the customization still satisfies a distinct requirement, it should remain.

The supplied exam source explicitly identifies D as the intended answer. Relevant topics: Agent Construction, SDK capabilities, abstraction boundaries, technical-debt management, retries, error handling, and agent state management.

===============

Option D treats prompts as production configuration and executable application behavior rather than informal text. A prompt modification can change output format, tool-selection behavior, policy enforcement, downstream compatibility, token usage, or model quality just as materially as an ordinary source-code change. Consequently, prompt changes need traceability, review, testing, and rollback.

A single source of truth prevents different application paths from silently using inconsistent prompt versions. Version control provides an immutable history showing who changed the prompt, what changed, and which application release used it. Pull-request review enables peer assessment and automated evaluation before deployment. The same configuration-as-code principle appears throughout Claude Code’s project configuration model, where shared project configuration is designed to live with version-controlled project artifacts.

A removes governance and creates an uncontrolled external state. B increases hidden runtime behavior and bypasses normal review. C solves fragmentation but leaves the exact governance failure described in the scenario—the ability to ship prompt changes without review.

Therefore, D provides reproducibility and controlled change management. Relevant Study Guide topics: prompt versioning, configuration management, source control, change review, reproducibility, regression testing, and rollback.

===============

D applies configuration as code and is the answer explicitly selected in the supplied examination source. Claude Code treats project configuration as part of the repository's working context. Anthropic documents project-level CLAUDE.md as the mechanism for providing project instructions and context, while shared project configuration belongs in .claude/settings.json. Crucially, Anthropic states that .claude/settings.json is intended to be checked into source control and shared with the team .

Keeping these files alongside the code establishes a single versioned relationship between application behavior and the configuration that influences Claude Code. Standard pull requests then provide diffs, peer review, ownership controls, CI validation, audit history, and straightforward rollback. A particular commit can therefore reproduce both source code and its associated Claude project configuration.

A creates competing copies and synchronization risk. B artificially separates behavior configuration from the code version it affects. C creates the same problem specifically for settings.json.

Developer-local values remain distinct: .claude/settings.local.json exists specifically for settings that should not be committed.

Relevant Claude Developer topics: Confia Management, CLAUDE.md, settings.json, configuration as code, source control, pull-request review, reproducibility, auditability, and team-shared Claude Code configuration .

D is the only option that addresses all three identified production-readiness defects. The supplied Claude Developer item explicitly selects D. A functioning happy path is insufficient for a maintainable Claude integration.

Prompt and model selection are configuration concerns that will change as prompts are evaluated, model versions evolve, or environments require different behavior. They should therefore be separated from unrelated business logic rather than scattered as inline constants. Claude API failures must also be handled deliberately. Anthropic documents typed SDK exceptions and defined HTTP error categories, including invalid requests, authentication failures, rate limits, server errors, overload, and timeouts. The official SDKs additionally retry appropriate transient failures.

Tests are required to verify the integration boundary, including successful behavior, malformed or unexpected responses, API error handling, and critical user workflows. Approving code without those controls pushes known reliability debt directly into production.

A and B knowingly merge incomplete production behavior. C improves configuration and test coverage but leaves a known API-failure path unhandled.

Relevant Claude Developer topics: SW Eng Foundations, code review, separation of configuration, error handling, integration testing, API resilience, maintainability, and production readiness .

===============

The supplied examination page marks B . The scenario already demonstrates why model upgrades must be treated as evaluated software changes rather than automatic replacements: the new model improves one metric while introducing a regression in another.

Anthropic's official model-selection guidance recommends creating benchmark tests specific to the application's use case, testing models with the application's actual prompts and data, comparing response quality and edge-case performance, and weighing performance against operational tradeoffs. Therefore, the correct action is to adapt the multi-section system prompt to the new model's behavior and repeat the evaluation. Only after the formatting regression is eliminated—or reduced below an explicitly acceptable threshold—should the upgrade proceed.

A incorrectly assumes that an 8% reasoning improvement numerically compensates for a 3% malformed-output rate; these metrics measure different consequences and cannot simply be subtracted. C treats the known incompatibility only downstream instead of first correcting the prompt/model interaction. D permanently rejects future improvement and is inconsistent with controlled lifecycle evolution.

The engineering principle is migration through regression testing and adaptation , not blind upgrading or permanent version avoidance.

Relevant Claude Developer topics: Systems Life Cycle, model migration, regression evaluation, prompt adaptation, compatibility testing, deployment gates, and continuous evolution .

===============

Option C correctly identifies threat categories introduced or significantly amplified by LLM-based application architecture. Prompt injection attempts to manipulate Claude into following adversarial instructions. Jailbreaks seek to circumvent behavioral or application safeguards. Data leakage can expose confidential information contained in system prompts, retrieved context, tool results, or conversation state. Unsafe model output becomes particularly serious when output is consumed by downstream systems or translated into tool actions.

Anthropic's official guardrail documentation explicitly distinguishes jailbreaks and direct prompt injection from indirect prompt injection. It recommends input screening, hardened system prompts, structured handling of untrusted tool content, least-privilege access, output screening, and continuous monitoring. Anthropic also warns that sensitive prompt/context data can leak and recommends post-processing and output filtering where required.

Traditional threats such as SQL injection or XSS remain relevant to the surrounding application, but they are not the additional AI-specific categories requested. B similarly describes conventional infrastructure threats. D is far too narrow because the SDK is only one component of the attack surface.

The supplied exam source explicitly marks C. Relevant topics: Claude App Design, threat modeling, prompt injection, jailbreaks, data leakage, output safety, least privilege, and defense in depth.

===============

Option A is the technically verified answer. Anthropic documents that multi-turn agentic requests resend the growing conversation context on subsequent turns. Prompt caching allows repeated prompt prefixes—such as system instructions, tool definitions, and prior unchanged conversation content—to be reused at the lower cache-read rate rather than repeatedly charged as ordinary uncached input. Anthropic specifically identifies caching repeated context as a major production cost optimization.

B is incorrect even though the supplied PDF contains a duplicated screenshot in which B appears selected. The Batch API is designed for asynchronous workloads that can tolerate delayed completion; it is not the appropriate mechanism for every interactive turn in a multi-turn conversation. Because the user requested verified answers rather than blindly reproducing marked selections, A is retained.

C discards potentially essential conversation state arbitrarily. D is a valid context-management technique in some long-running workflows, but summarizing after every turn creates additional model work and loses detail; it does not exploit repeated-prefix caching.

Therefore, A directly addresses the stated cost pattern while preserving conversation fidelity. Relevant Study Guide topics: Messages API, stateless conversation history, prompt caching, cached input tokens, multi-turn applications, token economics, and cost optimization.

===============

Option C is correct. Claude Code uses CLAUDE.md as the standard mechanism for supplying persistent project-level instructions and context. Anthropic's Claude Code documentation states that project instructions can be stored in ./CLAUDE.md or ./.claude/CLAUDE.md and shared with team members through source control. Appropriate content includes coding standards, architectural decisions, project conventions, build/test commands, preferred workflows, and information developers would otherwise need to repeat to Claude during every session. The /init command can also generate an initial CLAUDE.md based on the repository.

A wiki may be valuable for human documentation, but Claude Code does not automatically receive that material as project context. B creates developer-specific configuration and risks inconsistency across the team. D places the information in a human-facing README but does not use Claude Code's purpose-built persistent instruction mechanism.

Therefore, C gives the repository a shared, version-controlled source of Claude-specific project guidance. Relevant Study Guide topics: CLAUDE.md, project configuration, persistent instructions, repository context, team standards, and configuration management.

===============

The supplied Claude Developer item selects A . The symptoms—duplicated logic, tangled control flow, and increasing difficulty adding features—indicate accumulated technical debt in a component that has become structurally difficult to change. Because the problem is already materially reducing development velocity, deliberate refactoring should occur before additional feature pressure compounds it.

The objective is behavioral preservation with structural improvement. Duplicated prompt construction should be consolidated where the behavior is genuinely shared. Prompt preparation, configuration, API invocation, validation, error handling, and post-processing should have clear responsibilities. Control flow should be simplified so each stage is testable independently. Existing tests should first capture important current behavior so refactoring can proceed without silently changing application semantics.

B can be appropriate for minor cleanup, but the question describes broad systemic complexity. Mixing substantial structural refactoring into unrelated feature tickets makes scope, review, and regression analysis harder. C knowingly allows the debt to compound. D centralizes complexity rather than removing it and violates separation of concerns.

Relevant Claude Developer topics: SW Eng Foundations, refactoring, technical debt, DRY, separation of concerns, maintainability, modular design, regression testing, and control-flow simplification .

===============

Option B is correct because Claude API cost is fundamentally driven by usage volume multiplied by the token economics of the selected model and features. A credible projection must therefore estimate requests, average input tokens per request, average output tokens per request, applicable input/output prices, and expected prompt-cache behavior. Anthropic’s pricing documentation treats input, output, cache writes, and cache reads as distinct billable categories, with cache hits priced below standard input processing.

A practical forecast can be modeled as: request volume × expected per-request input cost plus request volume × expected per-request output cost, adjusted for cache-write/read rates and any other applicable pricing modifiers. The model should also include ranges for variance rather than only one point estimate, because token lengths and cache-hit rates will fluctuate in production.

Option A assumes a historical application has the same token profile and pricing, which may be false. Option C omits output tokens even though output pricing can be a material share of cost. Option D postpones the very token estimates the product team needs before launch.

Therefore, B is the complete projection method. Relevant Study Guide topics: token accounting, model pricing, prompt caching, cost forecasting, workload sizing, and production economics.

===============

Option B is correct because destructive production operations require deterministic enforcement outside the model’s probabilistic reasoning. Claude Code hooks can intercept lifecycle events before tool execution and explicitly allow, deny, or request further handling based on concrete rules.

Anthropic’s hooks documentation provides this exact security pattern. A PreToolUse hook can inspect a proposed command before execution and return a blocking decision. Anthropic’s example demonstrates blocking destructive operations such as drop table, while other commands proceed normally.

That mechanism can be adapted to database controls: block DROP, destructive DELETE, unauthorized schema modifications, or production writes; require explicit approval for high-risk operations; and allow read-only or known-safe queries automatically.

A merely increases the probability that someone might notice an unsafe operation and does not prevent execution. C assumes model capability can replace access controls, which is an unacceptable safety boundary. D is useful behavioral guidance but remains probabilistic and cannot guarantee prevention.

Therefore, B creates a deterministic control between model intent and side-effect execution. Relevant Study Guide topics: Claude hooks, PreToolUse, tool governance, deterministic enforcement, approval gates, least privilege, and destructive-operation protection.

===============

Option A correctly treats the MCP server as a reusable integration boundary rather than application-specific code. Model Context Protocol separates capability providers from consuming Claude applications by exposing standardized tools, resources, and prompts through an MCP-compatible interface. A shared server therefore needs explicit capability definitions, an appropriate transport or communication pattern, and a deployment location reachable by its intended clients.

Anthropic’s MCP documentation distinguishes remote HTTP-based integrations from local/client-managed connections. For remotely shared services, the server must be reachable from the consuming environment; client-side MCP helpers additionally support broader MCP capabilities such as resources and prompts.

B unnecessarily couples the server design to the first consumer and encourages application-specific evolution of what should be a reusable service boundary. C prevents multi-application deployment because only local developer sessions could reach the service. D abandons MCP entirely and recreates duplicated integrations in every consuming application.

Thus, A provides the correct lifecycle: define the MCP contract, expose resources/tools/prompts appropriately, select the transport according to topology, deploy the service, and let multiple applications consume the standardized interface independently. Relevant Study Guide topics: MCP architecture, reusable services, tools, resources, prompts, transports, and deployment topology.

===============

Option A is correct because model selection requires measurable non-functional requirements, not an undefined objective such as “best.” Anthropic's model-selection guidance explicitly frames the decision around capabilities, speed, and cost and recommends identifying those requirements before narrowing the candidate models.

“Quality” must also be translated into task-specific success criteria—for example classification accuracy, extraction correctness, acceptable false-positive rate, or human preference. Latency may be constrained by an interactive SLA, and cost may be defined per request, per user, or per completed task. Once those thresholds exist, representative evaluations can compare candidate models using the actual prompts and data.

B uses a generic benchmark that may not correlate with the application's workload. C chooses a model tier without requirements or evidence. D optimizes only for raw capability and ignores latency and cost, potentially producing an unnecessarily expensive architecture.

Thus, requirements definition precedes technology selection. Relevant Study Guide topics: Understanding Reqs, functional versus non-functional requirements, quality thresholds, latency budgets, cost constraints, acceptance criteria, workload-specific evaluation, and model-selection tradeoffs.

===============

Option B establishes a controlled boundary between probabilistic model output and deterministic downstream code. When structured data is machine-consumed, malformed JSON must be recognized as an explicit application error rather than allowed to propagate into parsers, databases, or other services.

Anthropic’s Structured Outputs documentation identifies malformed JSON, missing fields, inconsistent types, and schema violations as exactly the kinds of failures that can break downstream systems when unconstrained output is used. Current Claude capabilities can constrain responses using JSON Schema, and SDK helpers can provide parsing and validation.

Even when an application uses an older or unconstrained generation path, it should parse against the expected schema, record validation failure, and enter a bounded recovery path such as retry, repair, fallback, or controlled rejection. A human review of every request is unscalable. C removes a useful structured interface instead of making it reliable. D performs uncontrolled blind retries and provides no schema-aware error handling or bounded fallback strategy.

Therefore, B gives the application explicit failure semantics and protects downstream systems. Relevant Study Guide topics: structured output, JSON validation, schema enforcement, retries, fallback handling, defensive parsing, and downstream reliability.

===============

Option B establishes a layered testing strategy rather than relying on a single testing granularity. Unit tests validate deterministic functions and isolated components quickly. Integration tests verify boundaries between application code and Claude-related components such as API clients, tool execution, parsing, persistence, and error handling. End-to-end tests then validate the most important user workflows across the complete application stack.

This separation is particularly useful for Claude applications because deterministic software failures and probabilistic model-quality failures should not be treated identically. Anthropic's evaluation guidance recommends defining specific, measurable success criteria and constructing representative test cases to determine whether model behavior meets those criteria. These evaluations complement conventional software tests rather than replacing them.

A focuses exclusively on test-driven development; TDD can be valuable but does not define all required test levels. C preserves the existing ad-hoc methodology rather than establishing a repeatable quality strategy. D overuses expensive and slower end-to-end tests while omitting the faster diagnostic value of unit and integration tests.

The supplied exam source marks B as correct. Relevant topics: SW Eng Foundations, test strategy, unit testing, integration testing, end-to-end testing, Claude evaluations, regression coverage, and production reliability.

===============

Option A is correct because sensitive-data protection must occur before prohibited information crosses the application's trust boundary. If account numbers, passwords, or other secrets must not be sent to Claude, asking Claude to “ignore” those values after transmission does not satisfy the requirement. The application must determine which data is permitted to leave its controlled environment and transform or reject prohibited content before constructing the Claude request.

This principle is consistent with Anthropic's defense-in-depth guidance. Anthropic recommends input screening, limiting Claude's access to sensitive information, and applying least privilege so the model receives only information necessary for the task.

B is fundamentally too late: the sensitive information has already been transmitted to the model context before the system instruction can influence behavior. C is worse because logging can create an additional store containing the sensitive information and only detects exposure after it occurs. D implements some boundary filtering but deliberately leaves coverage incomplete until production incidents reveal additional patterns.

Therefore, A requires the boundary and its permitted data flows to be defined first, then enforced comprehensively through filtering, redaction, tokenization, or rejection. Relevant Study Guide topics: application boundaries, data minimization, sensitive-data handling, filtering, redaction, least privilege, and privacy-preserving design.

===============

Option C is the correct application of context isolation. A specialized step should receive the minimum relevant information required for its own task rather than inheriting an ever-growing global transcript. This improves signal-to-noise ratio, limits accidental cross-task influence, controls token usage, and makes individual components easier to evaluate.

Anthropic’s context-engineering guidance explicitly identifies multi-agent architectures as a technique for long-horizon work. Specialized subagents can operate with their own context windows and return condensed results to an orchestrating agent rather than exposing every agent to every intermediate detail. This architecture protects each reasoning process from irrelevant history while allowing the overall system to preserve necessary state.

Options A and D represent the opposite approach: indiscriminately loading the full prior history. Larger context is not automatically better; Anthropic warns that excessive context can introduce context pollution and degrade retrieval or attention to important information. B also fails because a single global prompt does not isolate state or tailor the information available to each processing stage.

Therefore, C best implements scoped reasoning boundaries. Relevant Claude Developer topics are multi-agent architecture, subagents, context isolation, context engineering, orchestration, and long-running workflow design.

===============

Option A follows disciplined production debugging: diagnose the actual failure mode before changing architecture or prompts. An output containing unsupported information might indeed be hallucination, but similar symptoms can result from stale conversation state, incorrect retrieval, unexpected tool output, prompt injection, incorrect request construction, or mismatched model/configuration versions.

A production trace should capture the user input, system instructions, relevant conversation history, retrieved content, tool calls and results, model/version, request parameters, response, and identifiers necessary to compare successful and failing cases. This establishes whether the model invented a fact or whether that fact entered context through another path.

B changes the model before establishing causality. C may eventually be useful if the confirmed problem is insufficient grounding, but implementing RAG before diagnosis can hide rather than explain the defect. D similarly changes prompting before verifying that prompt behavior is responsible.

The engineering sequence should be observe, reproduce, classify the failure, form a hypothesis, apply a targeted correction, and validate the correction with evaluations. Relevant Study Guide topics: production troubleshooting, observability, tracing, hallucination analysis, prompt injection, context failures, regression diagnosis, and lifecycle monitoring.

===============

Option D reflects the correct architecture-selection principle: extension mechanisms should be selected according to what the capability must accomplish rather than familiarity, novelty, or implementation convenience. Anthropic explicitly distinguishes Claude extension mechanisms by purpose. Built-in tools cover common capabilities already supported by the platform. Custom tools expose application-specific callable operations through defined schemas. Skills package reusable knowledge, instructions, and workflows. MCP connects Claude to external systems, APIs, databases, and services through a standardized protocol.

The decision therefore depends on scope and integration boundaries. A reusable procedural workflow may belong in a Skill; access to an external enterprise system may warrant MCP; a narrowly application-specific operation can be a custom tool; and a built-in tool should generally be preferred when it already satisfies the requirement.

A makes prior team experience the architectural criterion rather than requirements. B assumes newer technology is inherently better. C optimizes implementation convenience without considering maintainability, interoperability, or reuse.

Therefore, D is the appropriate selection rule. Relevant Study Guide topics: built-in tools, custom tools, Skills, MCP, extension architecture, reuse boundaries, and capability selection.

===============

Option C satisfies both the enterprise routing requirement and sound application architecture. Claude is available through Amazon Bedrock, and Anthropic provides Bedrock-specific SDK integration rather than requiring applications to call api.anthropic.com directly. Current Anthropic documentation describes Claude in Amazon Bedrock as operating through AWS-managed infrastructure with AWS-native authentication, billing, and security boundaries. Newer Bedrock integrations use the Messages API shape, allowing substantial application logic to remain consistent across provider environments.

Anthropic SDKs also provide dedicated Bedrock clients—for example, Python includes AnthropicBedrockMantle for current Bedrock deployments. Keeping business logic separated from provider-specific authentication, endpoints, model identifiers, and transport configuration reduces migration and maintenance risk.

A violates architectural simplicity by duplicating every call unnecessarily. B knowingly violates the enterprise requirement until migration occurs. D directly ignores the contractual routing constraint and is therefore invalid regardless of technical feasibility.

The correct approach is to make Bedrock the configured inference provider while keeping higher-level application and agent behavior decoupled from provider-specific implementation details. Relevant Claude Developer topics are Claude API mechanics, cloud-provider integrations, Amazon Bedrock, SDK configuration, authentication boundaries, model invocation, and provider abstraction.

===============

Option B is correct because the official Anthropic Python SDK is the supported abstraction for calling the Claude REST API from Python. Anthropic documents both synchronous and asynchronous clients, standardized request/response objects, streaming support, error classes, timeouts, request IDs, and built-in retry behavior for common transient failures. Using those documented primitives reduces boilerplate and keeps integration behavior aligned with the API.

Option A is technically possible, but it unnecessarily reimplements authentication headers, serialization, error mapping, retries, timeouts, and response handling that the SDK already provides. Direct HTTP is appropriate only when there is a specific reason not to use the supported client. Option C introduces another provider’s abstraction and a translation layer that is unrelated to the stated requirement and increases compatibility risk. Option D invokes shell commands from Python, creating needless process-management, security, and error-handling complexity.

The correct engineering principle is to use the highest-level supported client that satisfies the application’s requirements while retaining access to lower-level HTTP controls when genuinely needed. Therefore, B is the appropriate Claude integration pattern. Relevant Study Guide topics: Python SDK, REST abstraction, authentication, retries, exceptions, response parsing, synchronous/asynchronous clients, and maintainable API integration.

===============

A is correct because the workflow is state-dependent, non-linear, and iterative . The source explicitly identifies the graph-based pattern as the intended architecture for this scenario. Each stage—fact extraction, eligibility evaluation, and decision-making—has its own completion criteria, and the process may need to move backward from eligibility evaluation to fact extraction when information is incomplete. A graph representation naturally models these conditional transitions and loops.

Anthropic's current orchestration guidance supports workflows containing branching, loops, filtering, staged execution, and state-dependent control flow , rather than forcing every task through one fixed sequence. Dynamic workflow orchestration can use explicit control logic so the next processing stage depends on current state and previous results.

B concerns progressive output delivery, not workflow-state transitions. C provides a generic agentic tool loop but does not explicitly model distinct states or transition criteria. D is unsuitable because it prohibits the required return path between extraction and eligibility evaluation.

Therefore, a graph-based architecture provides the necessary conditional routing, iteration, and stage-specific validation.

Relevant Claude Developer topics: Agent Patterns, graph workflows, state transitions, conditional branching, loops, stage-specific success criteria, and agent orchestration .

===============

Option C applies the appropriate context-management technique. The issue is not simply whether the older conversation can physically fit into the context window; it is that stale details remain prominent enough to interfere with the model’s current task. Effective context engineering optimizes signal quality, not merely maximum token retention.

Anthropic’s context-editing documentation describes compaction as summarizing accumulated history and replacing the full history with a structured summary when context becomes large. This preserves important task state while substantially reducing low-value detail that can distract later reasoning.

A retains precisely the stale material producing the problem and therefore does not address context drift. B destroys all useful continuity between turns, including valid goals, decisions, and intermediate state. D is similarly excessive because limiting Claude to only the latest turn discards information that may still be required.

Compaction provides the correct middle ground: retain durable conclusions, current objectives, unresolved issues, and other high-signal state while compressing obsolete conversational detail. Relevant Study Guide topics: context engineering, compaction, conversation history, long-running agents, stale context, context drift, and high-signal state preservation.

===============

Option A applies the correct context-engineering strategy: remove stale tool results once they no longer contribute useful information to subsequent reasoning. Agentic workflows frequently accumulate search results, file contents, API responses, and intermediate artifacts. Keeping all of them indefinitely consumes the finite context window, raises token cost, and can reduce model focus by introducing low-value information.

Anthropic specifically documents tool result clearing for this situation. Context Editing can remove older tool results when the conversation grows, while preserving recent interactions and optionally retaining tools whose results must remain available. Anthropic describes old tool outputs such as retrieved files or search results as candidates for clearing after Claude has processed them.

Prompt caching in B solves a different problem: it can lower cost and latency for repeated static prompt prefixes, but cached tokens still constitute context and therefore do not eliminate context-window pressure. C changes model capability without solving the architectural cause. D maximizes context pollution.

The correct architecture is therefore to preserve high-signal state while pruning ephemeral intermediate outputs. This aligns with Claude Developer coverage of context engineering, long-running agents, context-window management, tool-result clearing, and efficient agent state management. Anthropic's broader context-engineering guidance likewise emphasizes curating the smallest high-signal context necessary for successful inference.

===============

Option D precisely matches Anthropic's SDK behavior. Official SDK documentation states that transient conditions are automatically retried two times by default with exponential backoff. The default retry set includes connection failures, HTTP 408 Request Timeout, 409 Conflict, 429 Rate Limit, and 5xx server errors. Developers can alter or disable this behavior through the SDK's retry configuration.

This directly eliminates A because the retry count is configurable. B is incorrect because rate-limit responses are among the conditions automatically retried by the supported SDKs; the API also provides retry-after information where applicable. C is incorrect because the SDK does not merely log these failures and continue without attempting recovery.

Application-level handling is still necessary after the configured retry policy is exhausted. Production code should distinguish permanent errors from transient failures, preserve request IDs for diagnostics, and decide whether higher-level fallback, queuing, or user-visible failure handling is appropriate.

Therefore, D correctly characterizes the SDK abstraction: useful retry defaults combined with developer control. Relevant Study Guide topics: Claude SDKs, transient errors, HTTP 429, retries, exponential backoff, configuration, and resilient API integration.

===============

Option B establishes the correct trust boundary. Fluent or confident language is not evidence that a generated claim is factually correct. If downstream systems treat output as authoritative data, the application must independently establish whether the output meets its correctness requirements before accepting it.

Validation can take several forms depending on the workload: compare generated facts against authoritative records, require citations or source references, constrain output to retrieved evidence, apply deterministic business rules, or use separate evaluation/classification stages. Anthropic's agent engineering guidance repeatedly emphasizes explicit evaluation criteria and validation rather than relying on apparent confidence.

A confuses sampling behavior with factual reliability. Lowering temperature does not establish factual correctness and may only make an incorrect answer more repeatable. C supplies maximum oversight but is unnecessarily expensive and removes useful automation even for low-risk, easily validated cases. D communicates uncertainty to users but does not protect downstream systems that automatically consume the response.

Therefore, B treats model output as untrusted until verified to the level required by the application. Relevant Study Guide topics: output validation, grounding, factuality, confidence calibration, source verification, trust boundaries, and downstream safety.

===============