Skip to content

6. Agent SDK

Domain 1 (27%) — Agentic Architecture & Orchestration — highest weight!

The agent loop — 5 steps — What happens inside every agentic turn 
1. Receive prompt → SDK yields SystemMessage (subtype: "init")
2. Claude evaluates → SDK yields AssistantMessage (text + tool calls)
3. Execute tools → SDK yields UserMessage (tool results)
4. Repeat steps 2–3 (each cycle = one TURN)
5. Claude produces text-only response
   → SDK yields final AssistantMessage + ResultMessage
ClaudeAgentOptions — cost and turn controls — How to cap cost and resuming sessions 
# Cap by turns (tool-use turns only)
options = ClaudeAgentOptions(max_turns=10)

# Cap by cost (good for production)
options = ClaudeAgentOptions(max_budget_usd=0.50)

# Resume a session
async for msg in query(
  prompt="continue",
  options=ClaudeAgentOptions(resume="session-name")
): ...
Escalation triggers — 3 conditions — When to hand off to a human — and common wrong answers
  1. Policy gap — request falls outside defined operating parameters
  2. Customer requests — explicit human escalation request
  3. Inability to progress — stuck, looping, or missing required information

Wrong answer traps on the exam:

  • ❌ "Use LLM self-reported confidence score" — poorly calibrated, never the correct answer
  • ❌ "Add sentiment detection to reduce escalation" — solves a different problem
  • ✅ Fix: explicit escalation criteria + few-shot examples first