Weight: 27%, the largest domain. About 16 of the 60 exam questions.
This domain covers how Claude agents run, how they coordinate, and how you make sure they stay within the rules.
| Topic | What it tests |
|---|---|
| 1.1 Agentic loops | Driving the loop with stop_reason, returning tool results, anti-patterns |
| 1.2 Multi-agent orchestration | Coordinator–subagent (hub-and-spoke) responsibilities |
| 1.3 Subagent invocation | Task tool, AgentDefinition, context passing, parallel spawning |
| 1.4 Workflow enforcement & handoff | Prerequisite gates, structured handoffs, multi-concern requests |
| 1.5 Agent SDK hooks | PreToolUse / PostToolUse interception and normalisation |
| 1.6 Task decomposition | Prompt chaining vs adaptive decomposition, per-file + integration passes |
| 1.7 Session state | --resume, --continue, forking, fresh start with a summary |
Mental model
- The model decides what to do next. Your code guarantees what must always or never happen.
- Subagents start with a blank slate: they know only what you put in their prompt.
- The coordinator owns decomposition, delegation, quality control and aggregation.
Exam heuristics
- If the scenario says something must happen always or never (compliance, money, identity), choose deterministic enforcement: a hook, a gate, or tool availability. Don't pick a stronger prompt.
- If a subagent's output is missing context, the cause is almost always that the coordinator didn't pass it explicitly.
- If a report has coverage gaps but every subagent did its job, look at how the coordinator decomposed the task.
Updated 2026-10-28