Augmented LLM
A single model call enhanced with retrieval, tools and memory. The basic building block of the taxonomy, and the right answer more often than its position at the bottom suggests: most problems need nothing more.
Exam context: Distractors inflate a multi-call coded pipeline into an "augmented LLM" or shrink it into one. Call count is the tell: an augmented LLM is one enhanced call.
See also: 1.1 Agent Architecture
Workflow
A system where LLM calls and tools are orchestrated through predefined code paths. The developer wrote the control flow. The model executes inside it, which is what makes runs predictable, testable and cheaper. Claude Code's dynamic workflows (as of September 2026) apply the same idea to multi-agent work: a script Claude writes decides what runs next and holds the intermediate results, so Claude's context holds only the final answer.
Exam context: A stem naming fixed, repeating, auditable steps keys the workflow, however capable the model in each step. A hand-coded chain is a workflow no matter how many model calls it contains.
See also: 1.1 Agent Architecture
Agent
A system in which the model directs its own process and tool usage, deciding each next step from environment feedback until the goal is met. Flexibility is the product. The price is latency, token spend, a wider error surface and a guardrail burden.
Exam context: An agent is justified when the path genuinely cannot be scripted: unknown step count, decisions dependent on intermediate findings. The extra spend must buy something a workflow could not deliver.
See also: 1.1 Agent Architecture
Context Isolation
The property that a subagent's context window starts fresh, with no parent conversation. Constraint and payoff at once: nothing reaches the worker except the delegation prompt, and nothing bulky returns to the lead except the worker's final message. As of September 2026 the one exception is a Claude Code fork, which inherits the whole conversation. Fork mode is off in the Agent SDK and -p runs unless you turn it on.
Exam context: The most-tested fact in the domain. Designs that assume workers see the orchestrator's conversation, or each other's findings when nothing passes them on, are broken by construction; the fix is always a fuller brief, never a respawn or a bigger model. Named subagents and agent teammates can message each other, but only through that explicit channel.
See also: 1.1 Agent Architecture
Delegation Prompt
The brief a subagent is spawned with: the only channel from parent to worker. A production brief carries an explicit objective, an output format, tool guidance, task boundaries, and every convention or decision the worker must honour.
Exam context: Multi-agent failures trace to the orchestrator first. Overlapping scopes produce duplicated sections, an unassigned topic produces a gap, and both are decomposition faults fixed in the briefs.
See also: 1.1 Agent Architecture
Agent Loop
The cycle every harness wraps: send the prompt with tool definitions; if the response requests tools, execute them, append the results, call again; stop when the model responds without tool calls or a cap trips. The raw Messages API imposes no loop semantics of its own.
Exam context: The loop's normal exit is a response containing no tool-use requests. There is no API-enforced iteration cap and no termination token. Limits exist only where the harness configures them.
See also: 1.2 Agent Construction with Claude
Turn
One round trip inside the agent loop: the model produces output including tool calls, the harness executes them, and the results feed back. The unit max_turns counts.
Exam context: Distractors redefine a turn as a token count, a session count or a subagent count. It is a tool-use round trip, nothing else.
See also: 1.2 Agent Construction with Claude
max_turns
The Claude Agent SDK option (maxTurns in TypeScript) capping tool-use round trips in a run. It defaults to no limit, and hitting it ends the run with the error_max_turns result subtype.
Exam context: The exact name is keyed. Invented-but-plausible near-misses (max_iterations, turn_cap) are the distractors, and a tripped cap is a configuration event, not a model failure.
See also: 1.2 Agent Construction with Claude
max_budget_usd
The Claude Agent SDK option (maxBudgetUsd in TypeScript) stopping a query when estimated spend reaches the set dollar value. Cumulative across the run, subagent spend included. The docs call it a good default for production agents.
Exam context: Context isolation does not mean cost isolation: subagent requests count toward total_cost_usd, spawns fail at the cap, and the run ends with error_max_budget_usd.
See also: 1.2 Agent Construction with Claude
ResultMessage
The final message of an Agent SDK run. Its subtype names the outcome: success, error_max_turns, error_max_budget_usd or error_during_execution (the docs list a fifth, error_max_structured_output_retries, as of August 2026). Every subtype carries total_cost_usd, num_turns and session_id. The result text is present only on success.
Exam context: Code that reads result without checking the subtype gets nothing useful on error runs. The documented recovery from a limit subtype is resuming the session with a raised cap, tightening the brief, or splitting the work.
See also: 1.2 Agent Construction with Claude
Harness
The code around the loop that production requires: turn caps, spend budgets, retries, tool-result validation, state persistence, observability. Execution guarantees live here; model behaviour lives in the prompt.
Exam context: The tested judgement is which side of the line a concern sits on. Prompted retries and prompted frugality are instructions the model may not follow; only the harness enforces.
See also: 1.2 Agent Construction with Claude
PreToolUse Hook
Code the harness runs before a requested tool call executes. It receives the tool name and input, and a deny permission decision blocks the call, with the reason returned to the model as the tool result.
Exam context: Prompts request, hooks enforce. The hook executes deterministically in the application process regardless of what the model was persuaded to attempt, which is what makes it a guardrail rather than guidance.
See also: 1.2 Agent Construction with Claude
Claude Managed Agents
Anthropic's hosted agent product, distinct from the Agent SDK: Anthropic runs the agent harness and the sandbox behind a REST API, providing stateful server-side sessions with minimal operational burden. In beta as of August 2026 (header managed-agents-2026-04-01), and the sandbox can be Anthropic-managed or self-hosted for data residency. As of September 2026, a self-hosted sandbox still sends tool inputs and outputs to Anthropic's control plane, and a session can be pinned to one agent version. Scheduled deployments run an agent on a cron schedule, and a coordinator agent delegates one level only, to a roster of at most 20 unique agents.
Exam context: Not "the SDK hosted for you". Items test who operates the harness in each deployment model: self-hosting keeps control and burden; Managed Agents trades control for minimal operations.
See also: 1.2 Agent Construction with Claude
Prompt Chaining
The workflow pattern that decomposes a task into a fixed sequence where each call processes the previous output, optionally with programmatic gates between steps.
Exam context: A gate is a one-way pass/fail checkpoint that stops bad output propagating; it never feeds critique back. A stem with a fixed sequence and checkpoints keys chaining, not evaluator-optimizer.
See also: 1.3 Agent Patterns and Frameworks
Routing
The workflow pattern that classifies an input and dispatches it to a specialised follow-up: a different prompt, tool set or model per category.
Exam context: The cue is distinct input categories deserving distinct treatment. The tempting distractor is one do-everything handler, or chaining, which fits one decomposable task rather than a stream of different ones.
See also: 1.3 Agent Patterns and Frameworks
Parallelization
The workflow pattern that runs LLM work simultaneously, in two forms: sectioning, where independent subtasks run side by side, and voting, where the same task runs several times for diverse or confident answers.
Exam context: Sectioning requires subtasks known and independent up front. Voting aggregates independent attempts in code, with no information flowing between them. It never converges on anything.
See also: 1.3 Agent Patterns and Frameworks
Orchestrator-Workers
The workflow pattern where a central model breaks the task down at runtime and delegates to workers. The pattern behind subagent architectures; its signature is that the subtasks are not known in advance.
Exam context: Sectioning's favourite distractor and vice versa. The dividing line is who decides the subtasks and when: enumerable up front means sectioning; discovered at runtime by the model means orchestrator-workers.
See also: 1.3 Agent Patterns and Frameworks
Evaluator-Optimizer
The workflow pattern pairing a generator with an evaluator that scores output against explicit criteria and feeds the critique back until the result passes, usually under a hard round cap.
Exam context: Keyed by explicit criteria plus iterative, critique-driven refinement. Voting is the adjacent distractor (no feedback between attempts) and the gate is the other (pass/fail without regeneration).
See also: 1.3 Agent Patterns and Frameworks
Memory Tool
Anthropic's cross-session persistence mechanism, generally available on the Messages API as tool type memory_20250818. Claude gets create, read, update and delete operations over files in a /memories directory that the application executes client-side. The application's handler carries the safeguards: it must reject paths outside /memories, and the docs suggest capping how much a view returns, with view_range for the rest.
Exam context: An agent remembers nothing between sessions unless the design adds a mechanism. Distractors rely on automatic recall, longer conversations or bigger windows; the keyed answer persists state explicitly.
See also: 1.3 Agent Patterns and Frameworks
Context Rot
The degradation of recall as a context window's token count grows, named in Anthropic's context-engineering guidance. The reason long-running agents need a design answer rather than a bigger window.
Exam context: The documented answers are compaction (the Agent SDK performs it automatically), structured note-taking, and sub-agent architectures. On the Messages API, as of September 2026, compaction comes in on-demand and token-threshold forms, and the docs recommend on-demand wherever it is available. Domain 1 keys the design move; the window-management technique detail belongs to Domain 6.
See also: 1.3 Agent Patterns and Frameworks
LangGraph
A low-level orchestration framework and runtime for long-running, stateful agents, from LangChain Inc but usable without LangChain. Control flow is modelled explicitly as a graph: deterministic and model-driven steps mixed, durable resumable execution, inspectable state.
Exam context: Keyed by needs that lead with graphs, statefulness and inspection. Recommending it for typed output validation is the classic identity swap.
See also: 1.3 Agent Patterns and Frameworks
Pydantic AI
The typed Python agent SDK from the Pydantic team: give the agent an output type and every run comes back validated against it.
Exam context: Keyed by schema-validated, typed outputs in Python. Graph-shaped control flow is LangGraph's identity, not this one's, and TypeScript is not its home.
See also: 1.3 Agent Patterns and Frameworks
Strands Agents
An open-source agent SDK for Python and TypeScript, built from production systems inside Amazon, taking a model-driven approach: the model drives the loop and orchestration code stays minimal.
Exam context: Keyed by a model-driven loop with minimal orchestration, often with an AWS-ecosystem cue in the stem. The identity swap to watch: schema-validated outputs are Pydantic AI's ground, not Strands'.
See also: 1.3 Agent Patterns and Frameworks