Domain 1 · 27% of the exam · ~16 of 60 questions
Agentic Architecture & Orchestration
Agentic loops, coordinator-subagent orchestration, context passing, workflow enforcement, hooks, task decomposition, and session state.
What this domain covers
This is the heaviest domain on the exam and the one whose ideas leak into every other. It is about the shape of an agent rather than the wording of a prompt: what drives the loop, who decides when it stops, how work is split across subagents, and what has to be written down so a crashed run can pick up where it left off. Questions here hand you a production system that is already misbehaving — a loop that never terminates, a subagent that has lost the thread, a workflow whose steps run out of order — and ask which structural change fixes it.
What the exam asks here
Drawn from the in-scope topic list in the official exam guide.
- Agentic loop implementation: control flow driven by stop_reason, tool result handling, and termination conditions that are not just an iteration cap
- Multi-agent orchestration: coordinator-subagent patterns, parallel subagent execution, and iterative refinement loops
- Subagent context management: explicit context passing, structured state persistence, and crash recovery from manifests
- Agent SDK hooks for intercepting tool calls and normalising data before it reaches the model
- Session state: resumption, forking, named sessions, and keeping session context isolated
The 7 task statements in this domain
- 1.1Agentic LoopsDesign and implement agentic loops for autonomous task execution
- 1.2Multi-Agent OrchestrationOrchestrate multi-agent systems with coordinator-subagent patterns
- 1.3Subagent Invocation and Context PassingConfigure subagent invocation, context passing, and spawning
- 1.4Workflow Enforcement and HandoffImplement multi-step workflows with enforcement and handoff patterns
- 1.5Agent SDK HooksApply Agent SDK hooks for tool call interception and data normalization
- 1.6Task Decomposition StrategiesDesign task decomposition strategies for complex workflows
- 1.7Session State and ResumptionManage session state, resumption, and forking
Common questions about Domain 1
How many questions on the CCAR-F exam come from Agentic Architecture & Orchestration?
At 27% it is the heaviest of the five domains, so roughly 16 of the 60 items. No other domain is weighted above 20%, which makes this the single best place to spend your first block of study time.
What ends an agentic loop?
The loop is driven by stop_reason on each API response, not by a fixed iteration count. An iteration cap is a safety bound that stops a runaway; the actual termination condition is the model returning a stop_reason that says no further tool use is requested. Confusing the two is one of the most reliably tested distinctions in this domain.
When should work be split across subagents rather than handled in one loop?
When the subtasks are genuinely independent and each needs its own context, so running them in parallel saves both wall-clock time and context budget. A coordinator delegates, collects structured results, and keeps the shared context clean. Splitting work that is sequential and interdependent just adds handoff points where state can be lost.
What has to be persisted for an agent run to survive a crash?
Enough structured state — typically a manifest of completed steps and their outputs — that a resumed run can tell what was finished from what was merely started. Conversation history alone is not sufficient, because replaying it re-executes side effects.