Agentic Architecture — Glossary

Agentic Architecture & Orchestration · 27% of the exam. 10 terms.

The model calls a tool, the client executes it and returns the result, and the cycle repeats until the model stops requesting tools.

Exam context: The only correct termination signal is stop_reason — not parsing the assistant's text for phrases like "I'm done."

The Messages API field explaining why generation stopped: end_turn, tool_use, max_tokens, or stop_sequence.

Exam context: Exam questions test whether you branch loop control on this field or on something fragile like text content.

A lead agent decomposes a task, dispatches subtasks to worker subagents, then synthesizes their results into one output.

Exam context: Pick this pattern when subtasks aren't fully predictable in advance and need dynamic decomposition.

Subagents report only to the orchestrator and never communicate with each other directly.

Exam context: A distractor answer that has subagents messaging each other peer-to-peer is always wrong for this exam's model.

A fixed sequence of LLM calls where each step's output feeds directly into the next step's input.

Exam context: Correct when a task decomposes cleanly into ordered subtasks — not when subtasks are independent (that's parallelization).

Classify the input first, then send it down one of several specialized prompt or model paths.

Exam context: Use when inputs fall into distinct categories that genuinely need different handling, not just different phrasing.

Run independent subtasks concurrently (sectioning) or run the same task multiple times for consensus (voting).

Exam context: Sectioning splits work; voting improves confidence on a single judgment call — know which the scenario calls for.

One call generates a response, a second evaluates it against explicit criteria and returns feedback for revision, looped until it passes.

Exam context: Exam scenarios use this when quality criteria are clear but generation quality alone isn't reliable enough.

Giving each subagent only the context it needs for its scoped task, not the orchestrator's full history.

Exam context: Protects both token cost and focus — a subagent drowning in irrelevant context is a common exam anti-pattern.

The parent spins up a subagent with a scoped prompt and toolset; it returns a final result, not a full transcript.

Exam context: Watch for options that have the orchestrator forwarding raw subagent transcripts back into its own context — usually wrong.