Option A is correct because the Message Batches API is designed for asynchronous, high-volume processing where results do not need to be returned interactively. Anthropic's API reference states that a Message Batch can contain many independent Messages requests and may take up to 24 hours to complete. That makes it appropriate for 10,000 overnight email-extraction jobs with a several-hour completion window.
Streaming in B solves a different requirement: it exposes partial response events while a single request is being generated, which is valuable for interactive user experiences or long-running synchronous requests, but it does not provide the workload-management advantages of a batch job. C processes items sequentially and unnecessarily sacrifices throughput. D can increase throughput with concurrent real-time calls, but it adds concurrency management and rate-limit pressure when the workload explicitly tolerates asynchronous completion.
The batch design also lets each request carry a custom identifier so results can be matched back to source emails even if completion order differs. Therefore, A is the intended Claude API choice. Relevant Study Guide topics: Message Batches API, asynchronous processing, high-volume workloads, request correlation, throughput, and non-interactive application design.

The supplied Claude Developer source explicitly marks A . The scenario contains the defining elements of an orchestrator/subagent architecture : decomposition of a larger objective, delegation of independent subtasks to specialized workers, and aggregation of their outputs by a coordinating agent.
This architecture is appropriate when subtasks can be performed with focused context or specialized tools.
Instead of forcing one agent to carry every intermediate detail, the orchestrator can formulate assignments, launch suitable subagents, receive condensed results, identify missing information, and synthesize the final research product. This also enables context isolation and potentially parallel execution.
Anthropic's published multi-agent architecture uses this orchestrator-worker approach for research workloads:
a lead agent decomposes a query, delegates work to specialized subagents, and then integrates their findings.
This is particularly useful where exploration is broad and individual subtasks benefit from independent context windows. The broader model-selection guidance also identifies orchestrator strategies as useful where work can be partitioned among worker models.
B describes storage rather than delegation. C is a context-management technique, not a decomposition pattern.
D centralizes all responsibilities in a single loop.
Relevant Claude Developer topics: Agent Patterns, orchestration, subagents, task decomposition, specialization, delegation, context isolation, and result consolidation .

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 A is the appropriate trust-boundary treatment for external content. Untrusted text can contain malformed data, adversarial instructions, prompt-injection attempts, or content deliberately constructed to alter the agent's behavior. It should therefore be validated or screened before inclusion and clearly represented as untrusted data rather than authoritative application instructions.
Anthropic's prompt-injection guidance distinguishes direct attacks from indirect prompt injection, where Claude processes third-party content such as webpages, emails, documents, or tool output containing hostile instructions. Anthropic recommends input validation and screening, least-privilege access, safe treatment of untrusted tool content, and screening content before Claude acts on it. Importantly, Anthropic notes that tool- result content is treated as untrusted data rather than as a reliable place for application instructions.
B erases the trust distinction and exposes the application to instruction/data confusion. C introduces a deliberately weaker security boundary and does not sanitize the data. D prevents legitimate use cases unnecessarily; untrusted does not mean unusable-it means the data must be handled defensively.
Therefore, A correctly combines validation with explicit trust separation. Relevant Study Guide topics:
prompt injection, untrusted content, input validation, sanitization, data/instruction separation, 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.