CCA-F slide notes (Domain 1 focus)

Rewritten notes from certification slide decks (lessons 2–3). For hands-on code, see Study map.

Multi-agent architecture

• Coordinator pattern: hub orchestrator delegates to workers; workers return structured summaries, not full tool logs.
• Single-agent limits: context overflow, sequential bottleneck, lack of specialization.
• Task tool: coordinators need "Task" in allowedTools; standard workers should not spawn uncontrolled sub-agents.

Sessions and context

• API is stateless — send full message history each turn.
• On --resume, tell Claude what changed on disk since the last session.
• Pass structured findings (objects), not prose dumps, between agents.
• sessions.fork: parallel exploration from one baseline without polluting the main thread.

Agentic loop (exam-critical)

• Exit when stop_reason is end_turn — not when you guess from natural language.
• On tool_use: append assistant tool blocks, then user tool_result messages, and continue.
• Anti-patterns: MAX_ITERATIONS as primary stop; parsing assistant text to decide completion.

Related pages

• 02 — Domain traps
• 03 — Exam scenarios
• Prep hub