Agentic Architecture — Quick Reference
Agentic Architecture & Orchestration · 27% of the exam. Density over narrative — this is for review, not first-pass learning.
Loop control signal
Branch only on stop_reason (tool_use vs end_turn). Never parse assistant text for intent, and never use an iteration counter as the primary termination mechanism — it's a safety cap, not the control flow.
Choosing an orchestration pattern
Ordered, predictable subtasks → prompt chaining. Distinct input categories → routing. Independent subtasks or need for consensus → parallelization. Unpredictable decomposition → orchestrator-workers. Clear quality bar, generation alone unreliable → evaluator-optimizer.
Subagent communication
Hub-and-spoke only. Subagents talk to the orchestrator, never to each other. A subagent returns a final result, not its full transcript.
Context to give a subagent
The minimum needed for its scoped task — not the orchestrator's full history. Isolation protects focus and token cost.
Default failure mode to avoid
Treating premature termination (stopping before tool_use resolves) and infinite fallback loops (never respecting a real end_turn) as equally likely exam traps — both come from not trusting stop_reason alone.