Option A matches the standard coordinator-managed subagent model. The coordinator invokes the web-search agent, receives its final result, selects the useful URLs or source records, and then supplies them explicitly when invoking the document-analysis agent. This keeps task ownership, provenance, and execution order visible to the coordinator.

The Claude Agent SDK subagent documentation states that a subagent begins with a fresh context window and does not inherit the parent’s conversation history or previous tool results. The information passed through the Agent tool’s prompt is the primary parent-to-subagent context channel. Consequently, the analysis prompt must include the source URLs, documents, retrieval notes, and any questions the analyzer must answer.

Option B requires nested delegation that is neither stated nor necessary and reduces centralized observability. Options C and D describe valid custom architectures only if the application has deliberately implemented a queue or shared memory system; neither is automatic Agent SDK behavior. Coordinator-mediated handoff is therefore the expected approach. For reliability, the handoff should use structured source records rather than an informal prose statement, ensuring that the analyzer receives identifiers, provenance, relevant excerpts, and analysis objectives.

Option C matches the distinction between reasoning about delegation and executing it. Defining subagents makes their descriptions available for selection, but the coordinator must still invoke the SDK’s subagent-spawning tool. Current Claude Agent SDK documentation calls this the Agent tool; Task was its earlier name and remains relevant to older SDK configurations. Anthropic’s Subagents in the SDK documentation instructs developers to include Agent in allowedTools so subagent invocations are approved automatically. Without that permission, an invocation can fall through to a permission callback or be denied under a non-interactive permission mode. Option A is unlikely because the configured subagent descriptions already tell Claude when each agent should be selected, although explicit prompting can improve invocation reliability. Option B misstates context isolation: context must be included in the spawning prompt, but that issue occurs after an invocation is attempted and does not explain the absence of all subagent executions. Option D would normally produce truncation evidence or incomplete output rather than consistent verbal promises with no tool call. The configuration should therefore permit Agent, explicitly request delegation where necessary, and log subagent invocation events.

Option C gives the reviewer access to the evidence currently missing from its prompt. An agentic review can use Grep, Glob, Read, language-server tools, and test commands to locate callers, follow imports, inspect type definitions, and verify whether a suspected compatibility issue actually exists. A turn limit controls cost while still allowing targeted exploration.

Anthropic’s context-engineering guidance recommends just-in-time retrieval: agents should retain lightweight references and dynamically load the information needed for the current task instead of preloading a large fixed context. Anthropic has also reported that code-review performance improves when the necessary repositories are available for gathering complete context. Option A may load many irrelevant files while still missing reflection, generated interfaces, dependency injection, or indirect call paths beyond the arbitrary two-hop boundary. Option B cannot reason reliably about code that the model cannot inspect. Option D improves coverage but fragments repository-wide relationships and may duplicate substantial work. An agentic loop can begin with the diff, identify uncertain cross-file effects, search only relevant paths, and report findings only after locating supporting code. Anthropic context-engineering guidance , Claude Code quality report

The reference to generic rendering patterns indicates that the current context has become saturated or that important earlier details are no longer receiving sufficient attention. Anthropic documents that performance can degrade as a Claude Code context window fills with conversation history, file contents, and command output; symptoms include forgetting earlier instructions and making less precise responses. ( https://code.claude.com/docs/en/best-practices )

Option C preserves the high-value rendering discoveries while isolating the physics investigation. The summary should identify the relevant rendering classes, synchronization boundaries, debug-overlay interfaces, and unresolved integration questions. A physics-focused subagent can then inspect collision and simulation code in its own context without adding dozens of additional files to the already crowded main conversation.

This explicit summary matters because an ordinary named subagent starts with a fresh, isolated context and does not automatically receive the parent’s full conversation history or previously read files. Claude constructs a delegation message containing the task context, so the essential rendering findings must be supplied deliberately. ( https://docs.anthropic.com/en/docs/claude-code/sub-agents )

Option A omits the integration context needed to guide the physics search. Option B destroys useful prior analysis. Option D keeps expanding an already degraded context and treats the symptom rather than controlling context growth.

Official references/topics: Context-Window Management; Subagent Isolation; Delegation Summaries; Codebase Exploration.

Option A uses the dedicated configuration mechanism for Anthropic’s managed Code Review service. The official Code Review documentation states that a root-level REVIEW.md is injected into every review agent as the highest-priority instruction block. It can redefine severity, suppress categories already enforced by CI, skip generated paths, and require concrete evidence before particular findings are reported.

The proposed rules map directly to the observed noise. Skipping lint and formatting findings eliminates duplication with deterministic CI checks. Excluding src/gen/* prevents comments on machine-generated code. Requiring a specific source line demonstrating incorrect behavior raises the verification threshold for rendering-helper findings without suppressing genuine defects.

Options B and C configure self-hosted GitHub Actions workflows, but the scenario concerns managed Code Review running through the Claude GitHub App on Anthropic’s infrastructure. Those workflow prompts do not govern the managed reviewer. Option D supplies useful general project context, but CLAUDE.md applies to all Claude Code tasks and has lower review-specific authority. The documentation states that REVIEW.md is the stronger control for changing what managed Code Review flags and how findings are reported. It therefore provides the most targeted solution with the least unnecessary infrastructure.

The correct objective is to construct an architectural map before consuming the full implementation. Identifying the base cache abstraction, its interface, and the classes that implement or invoke it gives the agent a dependency-guided path through the code. It can then inspect only the invalidation implementations and integration points relevant to the proposed trigger.

This approach protects the context window. Anthropic states that every file read occupies context and that model performance can deteriorate as the window fills. Its Claude Code guidance warns against unbounded investigation that reads large numbers of files and recommends narrowing the exploration or delegating it. ( https://code.claude.com/docs/en/best-practices )

Option A is too lexical: searching only for invalidate or expire can miss event-driven invalidation, overridden methods, cache-key mutation, and generic interface calls. Option B loads approximately 6,000 lines without first establishing relevance. Option C assumes that filename patterns and file size correlate with architectural importance; the largest files may contain incidental code while a small interface defines the entire design.

Option D follows control and type relationships rather than arbitrary file order. After reading the base class, the agent can search for subclasses, imports, construction sites, middleware hooks, and calls to the invalidation contract, progressively expanding only where evidence requires it.

Official references/topics: Context-Efficient Exploration; Dependency-Guided Reading; Architectural Interfaces; Narrowly Scoped Investigation

MCP allows a Claude application to connect to multiple servers and expose their enabled tools within the same agentic interaction. Anthropic’s MCP connector documentation explicitly supports connecting to multiple MCP servers in one request. Once connected, Claude can invoke a server’s tool when the user’s request corresponds to the capability described by that tool. ( https://docs.anthropic.com/en/docs/agents-and-tools/mcp-connector )

In this scenario, the git server can provide the branch-creation operation, the documentation server can locate the relevant links, and the Jira server can update ticket JIRA-123. The agent can coordinate these capabilities without requiring a separate conversational turn that restricts it to only one server. Tool descriptions and schemas tell Claude what each operation does and what inputs it requires.

Option B incorrectly describes sequential capability discovery. Server identity and tool definitions are established through the MCP configuration rather than discovered by querying every server for each operation. Option C is incorrect because connecting several servers does not inherently restrict the agent to a single selected server. Option D imposes a one-server-per-turn limitation that MCP does not require.

Access can still be controlled through allowlists, denylists, permissions, and per-tool configuration. Those controls restrict which tools are enabled; they do not change the fundamental ability to make tools from multiple configured servers available together.

Official references/topics: MCP Server Configuration, Multi-Server Tool Access, Tool Discovery, MCP Tool Permissions.

The pipeline must preserve source evidence while making inconsistencies explicit. Option D records the amount stated on the invoice separately from the total derived from extracted line items. A mismatch then becomes a machine-detectable validation condition rather than an invisible extraction defect.

This approach is superior because it does not silently overwrite source data or ask another model to guess which value is correct. Anthropic’s evaluation guidance recommends automated, code-based grading whenever the criterion can be expressed deterministically. Arithmetic reconciliation is precisely such a criterion. ( https://docs.anthropic.com/en/docs/build-with-claude/develop-tests ) In production, the summation should preferably be calculated by application code using normalized decimal values, even though the option describes the model populating calculated_total . The essential design principle remains the same: preserve stated_total , compute an independent total, compare them, and route discrepancies for adjudication.

Option A might improve behavior but cannot resolve genuine OCR corruption and could encourage the model to modify extracted values merely to create mathematical consistency. Option B introduces a second probabilistic judgment without new evidence. Option C is unacceptable for accounting data because it fabricates adjusted amounts and destroys fidelity to the invoice.

The schema should therefore expose both values and attach a validation status or review reason when they differ.

Official references/topics: Deterministic Validation; Human-in-the-Loop Review; Structured Output Design; Source-Fidelity Controls.

The implementation approach is unresolved and carries architectural consequences. WebSockets, Server-Sent Events, and polling differ in connection lifecycle, bidirectional communication, proxy compatibility, scaling requirements, deployment topology, reconnection behavior, and operational complexity. Claude should therefore inspect the existing application and infrastructure before modifying code.

Anthropic recommends separating exploration and planning from implementation when the approach is uncertain, the work affects multiple files, or the developer is unfamiliar with the relevant architecture. In plan mode, Claude can read files and analyze the system without making changes, produce a detailed proposal, and allow the developer or team to review the plan before execution. ( https://code.claude.com/docs/en/best-practices )

Option A commits to polling before determining whether its latency and load characteristics meet the requirement. Option B delegates a consequential architectural decision and immediate implementation to a single execution step without a review checkpoint. Option D prematurely selects the most infrastructure-sensitive option and assumes later refactoring will be inexpensive.

Option C creates the correct sequence: examine the frontend and backend architecture, document requirements, compare alternatives, identify infrastructure constraints, obtain approval, and only then implement the selected design with appropriate tests and observability.

Official references/topics: Plan mode, explore–plan–implement workflow, architectural trade-off analysis, review checkpoints.

Option D allows orchestration effort to scale with the actual request. A simple factual query may require only the web-search agent and a direct coordinator response, whereas a comparative investigation may require web research, document analysis, synthesis, and report generation. Anthropic’s Building Effective AI Agents describes the orchestrator-workers pattern as a central model dynamically identifying subtasks, delegating them, and combining the results. It is specifically appropriate when the required subtasks cannot be predicted reliably in advance. Anthropic’s multi-agent research architecture likewise emphasizes varying the number of agents and tool calls according to task complexity. Option A introduces an inflexible binary decision and still sends every non-factual request through the full pipeline. Option B requires labeled data, ongoing retraining, and reliable definitions of the “optimal” agent combination. Option C is easier to implement but will become brittle as new query types appear. A capable coordinator can examine the requested output, necessary evidence, source requirements, and analytical depth at runtime, then invoke only the specialists that materially contribute to the answer.

Option C preserves each commit’s existing model, prompt, and independent processing structure while applying Anthropic’s reduced batch pricing. The official Message Batches documentation states that batch processing reduces input and output token costs by 50%. Requests are processed independently and asynchronously, and each response can be correlated with its original commit through a unique custom_id.

This workload is a strong batch candidate because it contains many independent requests and does not require an immediate response. Most batches complete within one hour, although processing can continue for up to 24 hours. The pipeline must therefore tolerate asynchronous completion and retrieve the results when processing finishes.

Option A changes the request architecture, increases the risk of context or output-limit failures, and makes individual retry and result association more difficult. Reducing request count does not inherently reduce per-token pricing. Option B improves throughput but concurrency does not change token rates. Option D would reduce cost but violates the explicit requirement to retain the same model tier. The Message Batches API is the only option satisfying every stated constraint.

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

Option B introduces an evaluator-and-refinement loop at the correct orchestration layer. The coordinator already owns the research plan and delegation decisions, so it should inspect the synthesis result against the required questions, identify coverage gaps, and issue focused follow-up assignments. Anthropic’s description of its multi-agent research system follows this pattern: the lead agent synthesizes returned findings, determines whether additional research is required, and creates new subagents or refines its strategy before producing the final result. Increasing the initial query breadth, option A, may generate additional irrelevant material and cannot guarantee that unforeseen gaps will be covered. Option C merely documents the incompleteness instead of correcting it. Option D weakens role separation by giving the synthesis agent search capabilities, increasing tool complexity and bypassing the coordinator’s centralized tracking. Targeted re-delegation preserves specialized responsibilities and creates an observable sequence of research, evaluation, refinement, and resynthesis. The coordinator should also maintain explicit coverage criteria and limit the number of refinement rounds so the system improves completeness without entering an uncontrolled research loop.

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

The agent must preserve durable state for several concurrently open issues without keeping every conversational turn in the active context window. Option D creates a structured issue ledger containing identifiers, issue type, requested action, current status, completed steps, unresolved questions, and relevant tool outputs.

Anthropic recommends structured note-taking for long-horizon agents. In this pattern, critical state is written outside the context window and retrieved later, allowing the agent to preserve dependencies and progress across dozens of tool calls. Anthropic also describes compaction as retaining critical details while discarding redundant raw messages and historical tool output. ( https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents )

Option A is better than retaining the entire transcript, but a narrative summary can merge separate issues or obscure exact identifiers and statuses. Option B will discard the refund conversation because it occurred outside the most recent 30 turns. Option C may recover current backend facts, but it cannot reconstruct conversational commitments, prior explanations, or why a particular action remains pending.

A structured context layer should maintain one record per issue and be updated after every material action. The active prompt can include compact summaries of all open issues and expanded details only for the issue currently being discussed.

Official references/topics: Structured note-taking, external agent memory, context compaction, multi-issue state management.

Option C identifies the fundamental trade-off introduced by batch processing: lower cost in exchange for asynchronous completion and potentially substantial latency. Anthropic states that most Message Batches complete within one hour, but results may become available only when all requests finish or after 24 hours, whichever occurs first. Therefore, the workflow must remain useful even if security findings arrive considerably later than they would through synchronous requests.

The reviews are explicitly non-blocking, so batch processing can be suitable if developers can still act on delayed findings through follow-up commits. Option B describes an implementation requirement because batch results may be returned out of submission order, but Anthropic provides a deterministic solution through each request’s unique custom_id. Option A incorrectly focuses on near-instant feedback, which batch processing does not provide. Option D is not decisive because batches support independent Messages API requests containing system prompts, tool use, and multi-turn conversation content. The primary decision is therefore whether the maximum practical feedback delay is acceptable to the development workflow. Anthropic Message Batches documentation

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

A mid-process escalation should transfer the decision-ready state accumulated by the agent. The human reviewer needs the verified customer identity, relevant order information, the promotional-pricing discrepancy, the reason approval is required, and any actions already attempted. Option B preserves this information in a concise, structured handoff while avoiding unnecessary repetition of the complete raw transcript.

Anthropic’s tool-design guidance recommends returning high-signal information and stable identifiers containing only what Claude or the next workflow participant needs to determine the next action. Anthropic’s context-engineering guidance similarly advocates structured notes that preserve critical state and dependencies without retaining every redundant tool result. A structured escalation payload applies both principles and reduces handling time for the manager. ( https://platform.claude.com/docs/en/agents-and-tools/tool-use/define-tools )

Option A discards the investigation already completed. Option C violates the agent’s authorization boundary and risks an impermissible financial action. Option D provides auditability, but a reference ID alone forces the human to reconstruct the case from an excessively broad transcript. Human control must remain meaningful when an agent encounters a decision outside its authority; the agent should pause and hand the decision back with sufficient supporting context. ( https://www.anthropic.com/research/trustworthy-agents )

Official references/topics: Structured agent handoffs, high-signal tool results, human-control boundaries, persistent structured state

The error response must explicitly communicate both control semantics and user-facing meaning. Setting " retriable " : false tells the agent that repeating the same operation cannot change the outcome. Providing a customer-friendly explanation allows Claude to respond accurately without exposing internal implementation details or inventing its own interpretation of the business rule.

Anthropic recommends returning failed tool operations with an error indicator and sufficiently informative content so Claude can decide whether to retry, request another input, or explain the limitation. Tool interfaces should provide detailed descriptions and structured parameters rather than forcing Claude to infer operational behavior from ambiguous text. ( https://platform.claude.com/docs/en/agents-and-tools/tool-use/build-a-tool-using-agent?utm_source=chatgpt.com )

Option A appropriately limits automatic retries to technical failures, but it does not solve the stated customer-response problem unless business failures also carry clear semantics. Option B relies on fragile parsing of human-readable messages. Option C adds latency and another tool dependency, and eligibility may still change or fail for reasons not covered by the preliminary check.

A complete schema should identify the error code, category, retryability, customer-safe explanation, and permitted next actions. Technical errors can similarly return " retriable " : true with retry guidance and a maximum-attempt policy.

Official references/topics: Structured tool errors, retryability classification, customer-safe explanations, resilient MCP contracts.

Option B applies Anthropic’s dedicated asynchronous bulk-processing mechanism while preserving the existing model and prompt for every commit. The Message Batches API accepts independent Messages API requests, each identified by a unique custom_id, and charges both input and output usage at 50% of standard API prices. The approximately 12-hour latency allowance makes the release-notes workload well suited to batching because immediate results are unnecessary.

Option A may reduce wall-clock completion time, but concurrency does not alter the API’s per-token price. Option C changes the task structure and risks exceeding context or output limits, mixing commit-level classifications, complicating retries, and making it harder to associate errors with individual commits. A single large request also does not automatically consume fewer tokens because the model must still process all commit content. Option D violates the requirement to retain the same model tier and output-quality profile. Batch results may arrive in an order different from submission order, so the pipeline must associate every response with its original commit through custom_id. This provides lower cost without changing the individual review prompts. Anthropic Message Batches documentation

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

Option A exposes both independent invocations in the same assistant turn, allowing the Agent SDK or application tool runner to execute them concurrently. The coordinator can then receive both results together and continue with synthesis only after the independent research branches have completed.

Anthropic’s parallel tool-use documentation explains that a response may contain multiple tool-use blocks. Independent, read-only operations can be executed concurrently to reduce latency, after which all corresponding tool results should be returned together. The term “Agent” is used here because current Claude Agent SDK releases renamed the earlier “Task” tool.

Option B may shorten individual execution but does not eliminate the sequential waiting pattern and could reduce research quality. Option C expresses the desired behavior but does not correct an orchestration implementation that processes only one tool call per turn. Option D introduces unnecessary coordinators, duplicated context, and substantially more complex state management. A single coordinator issuing both independent Agent calls preserves centralized monitoring and result association while removing the avoidable serial dependency. The runtime must process every returned tool call concurrently rather than stopping after the first one.

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

Option C parallelizes independent work while retaining centralized orchestration. Each precedent can be analyzed without waiting for the previous precedent, so the coordinator can divide the 12 cases into balanced subsets and invoke several document-analysis subagents concurrently. It then receives their final outputs, records which precedents completed or failed, and aggregates the results before synthesis. Anthropic’s multi-agent research architecture uses an orchestrator-worker pattern in which the lead agent creates specialized subagents that operate in parallel and return findings for consolidation. Keeping spawning decisions at the coordinator also produces a clearer execution trace for monitoring and debugging. A generic asynchronous queue, option A, adds infrastructure but does not define how results remain associated with the correct research task. Options B and D create nested or recursive delegation, making execution paths, permissions, failures, and token consumption harder to observe. Anthropic also cautions that multi-agent systems consume substantially more tokens than ordinary interactions, so unbounded recursive decomposition is inefficient. Coordinator-controlled parallel fan-out followed by deterministic aggregation provides the latency improvement without sacrificing operational visibility.

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

Option C is the strongest choice among the listed approaches because it moves structured data generation from unconstrained response text into a schema-governed interface. Prompt-only JSON instructions, regular-expression extraction, and corrective retries are probabilistic recovery mechanisms: they may reduce failures, but none guarantees that the first response satisfies the contract. Anthropic now provides two schema-constrained mechanisms: JSON Structured Outputs through output_config.format and strict tool use through a tool definition with strict: true. Strict tool use uses grammar-constrained sampling so the tool input conforms to the declared JSON Schema, including required properties and data types. Therefore, the production implementation represented by option C should define the extraction tool’s input_schema and enable strict mode. The application reads the tool_use block as the extracted record; it does not need to scrape prose or search for braces. Option A remains vulnerable to preambles or malformed output. Option B can select a JSON-looking substring but cannot establish schema validity. Option D adds latency and cost and can still fail repeatedly. The schema-constrained tool contract is consequently the most reliable architecture available in these options.

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

Option B preserves the formal failure indicator while making the returned content operationally useful. Anthropic’s tool-use guidance states that when a tool fails, the application should return the error through tool_result content and mark it with is_error: true . Claude can then use that information to retry with corrected input, seek clarification, or explain the limitation. ( https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/bash-tool )

The two example messages identify both the cause and the appropriate recovery path. “Order not found” indicates that repeating the same identifier will not help and proposes an alternative lookup method. “Database timeout” identifies a transient infrastructure condition for which a retry is reasonable. This eliminates blind repetition without removing Claude’s ability to adapt.

Option A can be useful for tightly bounded low-level retries, but it does not solve permanent errors or tell the agent what happened after retries fail. Option C incorrectly disguises a failed operation as successful tool content. Option D moves semantic knowledge into hardcoded orchestration logic, making the system more rigid and duplicating information already known by the tool.

The most reliable tool interface combines the error flag, a stable error type, a retryability indicator, a concise causal explanation, and a recommended recovery action.

Official references/topics: is_error handling, instructive tool results, recovery guidance, resilient MCP tool design.