Key takeaways

• A hook is deterministic code that fires on a lifecycle event (PreToolUse, PostToolUse, Stop, PreCompact, SubagentStop), unlike a skill, which the model chooses. Hooks run in your process and cost no model context.
• PreToolUse hooks can veto a tool call, complementing Chapter 7's approval policy with custom rules.
• Observability comes from the session log (JSONL of every message, tool use, and result), which enables rewind, resume, fork, and audit.
• The Agent SDK exposes allowed_tools, max_turns, max_budget_usd, effort, and setting_sources, and returns a result subtype like error_max_turns. This makes budget-aware agents an engineering task, not a hope.
• Hooks plus budgets are what make an agent safe to run unattended; observability is what lets you debug and audit it.

Original sources: the Claude Code six-layers architecture breakdown and Anthropic's How Claude Code works.

---