Answer: B. The symptom described – instructions which contaminate each other between heterogeneous capacities – is the canonical signal of division. The orchestrator-router isolates each specialty: short and clean prompt, minimal tools, testability by agent. A worsens the contamination (even longer prompt). C (consensus) responds to a problem of reliability of the same task, not heterogeneity of tasks. D treats the symptom, not the cause: the power of the model does not replace the separation of responsibilities.

Answer: B. Context isolation is the central architectural benefit: the thousands of exploration tokens (pages read, dead ends, attempts) remain in the context of each sub-agent and never clutter that of the orchestrator, who reasons on dense syntheses. A is wrong — the total cost is often superior (4 contexts instead of one); we pay in tokens what we gain in quality and latency. This is false: a single agent can also call different models per request. D is false: parallelization does not provide any determinism.

Answer: B. The consensus multiplies the cost by N; it is only justified when the cost of an error greatly exceeds the additional cost of inference, typically high-stakes decisions with an element of judgment. Described workflow, not consensus. C and D are the opposite of reality: consensus increases latency and cost.

Answer: B. Routing is a classification: closed, high volume task, where latency and cost dominate. A simple call with constrained output on a fast model is the right tool. A is a classic exam anti-pattern: deploying an agentic loop where a call is enough. C throws the necessary semantic capacity (the tickets are in natural language). D is expensive and non-deterministic over-engineering.

Answer: B. The architectural principle: human friction is placed at the point of no return, proportionate to the risk (amount threshold). Destroyed the usefulness of the agent (each read validated by hand). It comes too late for irreversible action — the audit completes but does not replace approval. D confuses instruction and guarantee: a prompt is never a guarantee of execution; only the code (approval gate) is one.

Answer: C. Graceful degradation pattern: partial results preserved, targeted retry with backoff (backoff = increasing wait between attempts), and transparency on the gap if the failure persists — the summary says what it does not cover. A wastes 5 successes. B produces a falsely exhaustive synthesis: it is the worst choice in terms of reliability (invisible gap). D degrades the experience for a partial and recoverable failure.

Answer: B. The sustainable state lives out context (database, session store) in structured form; upon recovery, we reinject the minimum useful amount. A works but explodes in tokens and latency over the steps — and ends up saturating the window. This is a fundamental factual error: the API is stateless, the model does not “remember” anything between two calls. D externalizes to the user what the architecture must guarantee.

Answer: C. Decision rule: if the path is known in advance → workflow (deterministic, cheaper, testable); if the path depends on intermediate discoveries → agent. The breakdown diagnosis is the agent's textbook case: it is impossible to script the steps in advance. A, B and D are fixed-stage pipelines — deploying an agent there is over-engineering that is punished by review.

Answer: B. The tool description is the selection prompt: the model chooses based on descriptions, not implementations. Production quality descriptions include scope, contraindications (“do not use for…”) and distinctions between similar tools. It costs a lot to get around a specification defect. C reduces variance but not semantic confusion. D adds error surface instead of removing it.

Answer: B. The canonical separation: tools = actions that can be invoked by the model with parameters (search, write); resources = contents addressable by URI (URI = Uniform Resource Identifier) ​​that the application loads in the context, for reading; prompt = reusable interaction templates exposed to the user. D exactly reverse the first two. This is false: the three primitives have different control cycles (model / application / user).

Answer: B. Granularity is conceived from the point of view of agent task , not the structure of the underlying API: a tool = an intention. Consolidation reduces back and forth, tokens and opportunities for error. A does not solve the structural problem. C trades over-granularity against the risk of data injection and destruction. D destroys the value of the typed schema: the model no longer has any guidance on possible actions.

Answer: A. stdio (stdio = standard input/output, standard process input/output): local server, subprocess, minimal latency, security by local isolation — perfect for workstation tooling. Streamable HTTP: remote server, shared between clients, with authentication layer (e.g. OAuth) — perfect for the business. B is a fashion reflex, not an architect’s. This is false by definition of stdio. D is false: MCP primitives are transport independent. ⚠ The exact names of transports evolve with the MCP specification — check the current version.

Answer: B. Tool error is a reasoning data : returned in the result (error field or is_error ), formulated in an actionable way, it allows the agent to replan (propose room A). Transforms a recoverable business hazard into a technical failure. C lets the model hallucinate a success or a cause. D is the worst case: lying to the model guarantees a final confidently false answer.

Answer: B. CLAUDE.md is loaded automatically in the context at each session: each line costs tokens at each interaction. We put the operational contract that the agent cannot guess by reading the code — exact commands (make test-unit , not “run the tests”), conventions, dangerous areas. A and D dilute the signal and inflate the cost. C duplicates a document written for another audience with another level of detail.

Answer: B. Headless mode (-p / --print ) executes a single query without interaction; --output-format json makes the output reliably parsable by the pipeline; and in CI, permissions are declared explicitly (tool allowlist) rather than being granted en bloc. This is the recurring security trap: bypassing all permissions in an automated environment that can execute arbitrary code. ⚠ CLI flags are changing — check claude --help on the installed version.

Answer: B. Sorting criterion: what must be knew constantly goes in CLAUDE.md (loaded at each session); what is invoked on demand with arguments goes into a slash command (loaded only on call — zero cost the rest of the time). A, C and D are permanent knowledge → CLAUDE.md . B is a one-time parameterized ritual → slash command.

Answer: B. The plan mode separates the decision from the execution: the agent explores the code, proposes a complete plan, and the human arbiter Before all writing. Its value increases with the scale and irreversibility of the change (multi-files, migration). A, C, D are trivial and local modifications where the plan mode adds friction without benefit.

Answer: B. Cardinal distinction of the examination: CLAUDE.md = order (the model may forget it or bypass it); permissions configuration = guarantee (applied by the program, not by the model). Version settings.json in the repository makes the policy uniform and auditable for the entire team. Confused deposit and guarantee. C produces 12 divergent policies. D destroys the main use (running the tests) to avoid a finely manageable case.

Answer: B. The canonical pattern: event triggering, non-interactive mode, targeted context (the diff, not the entire repository), exploitable structured output, and read only — the magazine informs, the human decides. A does not scale. C crosses the red line of irreversible action without humans. D explodes the context and the cost for marginal gain: the diff and its surroundings are enough for the review.

Answer: B. Two API behaviors to know by heart: (1) the prefill text (prefilling of the start of the response) is not not repeated in the exit — the generation picks up after him; (2) the triggered shutdown sequence is not not included in the flipped text, and everything that followed is cut off; stop_reason worth "stop_sequence". Result: the application receives exactly the useful value, without tags or chatter. This is the canonical deterministic extraction pattern. ⚠ Volatile behavior: check current API documentation.

Answer: B. The CoT helps tasks where the answer results intermediate steps (A, C, D). On a verbatim copy, the “think” step invites the model to reformulate, standardize or paraphrase — exactly what we don’t want. Architect’s reflex: the CoT is a measurable tool, not a reflex; it is activated when the evaluation shows a gain.

Answer: B. The use of a tool as a structured output “mold” is the most restrictive mechanism offered by the API: the generation of arguments is guided by the schema (types, required fields, enumerations), and tool_choice forces the call. Downstream programmatic validation remains mandatory (belt + suspenders guarantee). A and C improve the probability without guarantee. D adds cost and latency and can introduce new errors. ⚠ Structured output mechanisms are evolving quickly — check common API options.

Answer: B. The few-shot examples teach decision boundaries , not the center of the distribution: the model already handles easy cases well. The clear-cut borderline cases show where the line passes; the case of refusal defines the perimeter. A and D waste the budget on what is already acquired. C leaves the composition to chance — the selection of examples is a design decision, not sampling.

Answer: B. Three long context best practices combined: large documents in high of the prompt, request in END (final instructions are best followed after extensive context), and citation anchoring (“cite relevant passages first”) forces the model to rely on the actual text before concluding. Reverses the optimal placement. C fragments without benefit. D is false: position in context has a measurable effect.

Answer: A. The four pillars of a harness: annotated and versioned reference set, metrics programmatic when the ground truth is objective (the extraction is: an amount is right or wrong), automation (the evaluation runs like a CI test), and historical (detecting regressions). The LLM-judge (B) reserves subjective qualities (tone, usefulness) — and calibrates itself against human judgments. C has no coverage or reproducibility. D measures everything except what matters.

Answer: B. Compaction is the standard pattern of long-running agents: context is a managed resource, not an infinite log. The summary retains the essential decision-making elements; raw details remain accessible out of context (files, database) if necessary. A loses information without control over which . C throws 2 hours of work. D is a caricature of A.

Answer: A. The cache works by exact prefix : any modification upstream of the cache point (a word, a tool, an order of blocks) changes the prefix, therefore invalidates the cache. B is precisely the intended use: the stable prefix is ​​reused, only the sequence varies. C describes the typical freshness window (≈5 minutes⚠, recharged on each access), not an invalidation. D does not invalidate: the cache is linked to the account/prefix, not to the machine. Architectural consequence: content stable in the lead (system, tools, documents), content variable in tail . ⚠ Volatile durations and prices.

Answer: B. Triple defense: exit door (allowing admission of ignorance reduces the pressure to invent), the anchoring (obligatory citations from sources), and the programmatic verification (a non-existent citation is detectable by code — this is a guarantee, not an instruction). Reduced the variance, not the invention: a hallucination at temperature 0 is simply reproducible. It is an instruction without a mechanism. D can make it worse (more noise, diluted context).

Answer: B. The complete resilience pattern: exponential backoff (immediate retries worsen congestion - A is a self-inflicted denial of service), jitter against the herd effect, ceiling to avoid looping endlessly, respect for server indications, and upstream traffic smoothing. This is a disproportionate strategic decision for a transitional and normal phenomenon. D confuses slowness with refusal: a 429 responds quickly, extending the timeout changes nothing.

Answer: B. Provenance requires being able to replay the causal chain: what (versioned) prompt, what model, what context, what tool actions, what output. Without the prompt version and entries, an incident is undiagnosable. Data protection applies to the logs themselves (PII — Personally Identifiable Information — minimized, encrypted, purged when due). A does not allow any causal diagnosis. This is ultimately illegal (unlimited retention of personal data) and unusable. D confuses bit by bit reproducibility and traceability: we can always trace what was provided and decided.