Agent SDK hooks

• PostToolUse hooks intercept tool results for transformation before the model processes them.
• Interception hooks examine outgoing tool calls and enforce rules, such as blocking refunds above a threshold.
• Hooks give deterministic guarantees; prompt instructions give probabilistic compliance.
• Normalising mixed formats (Unix timestamps, ISO 8601 strings, numeric status codes) from different MCP tools is a classic PostToolUse job.
• A blocking hook can redirect to an alternative workflow such as human escalation instead of just failing.

• Implement PostToolUse hooks that normalise heterogeneous data before the agent reasons over it.
• Implement interception hooks that block policy-violating actions (refunds over $500) and redirect to escalation.
• Choose hooks over prompt enforcement whenever a business rule needs guaranteed compliance.

• "Hooks are only for logging." They transform data and enforce policy.
• "A blocking hook can only fail the call." It can hand the agent a different path.
• "The model can be asked to convert formats itself." It can, but it will do so inconsistently and at token cost.

Policy thresholds enforced only in the prompt, and prompts bloated with format-conversion rules that a hook should handle.

• 1.4Workflow enforcement and handoff
• 2.4MCP servers in Claude Code and agents
• 5.1Conversation context management