Advanced Claude Code Hooks: Controlling Sub-Agent Behavior (ltscommerce.dev)

Anthropic’s Claude Code hooks can now be used to enforce fine-grained execution policies across parallel sub-agents — for example, preventing concurrently spawned agents from running unit test suites that share a SQLite database. The key technique is a PreToolUse hook (runs before any tool invocation) implemented in Python: it inspects the parent process ID (PPID) to detect sub-agent context (parent process named "claude"), uses regex pattern matching to classify commands (block phpunit, Infection, bin/qa -t unit; allow bin/qa -t allCs/allStatic), and returns a structured JSON error to block unsafe operations. Hooks are configured in Claude’s settings and are designed to be very lightweight and fail-open (allowing commands if detection fails) to avoid accidental disruption. For the AI/ML tooling workflow this matters because it preserves the productivity gains of parallel sub-agents while avoiding resource contention and nondeterministic failures (e.g., SQLite write locks, shared-fixture races, simultaneous schema changes). The pattern generalizes to migrations, file-system locks, API rate limits, and build artifacts. Best practices: make regexes specific, keep hooks fast, fail open for safety, and provide clear JSON explanations. This approach lets main agents coordinate risky operations while sub-agents run safe static checks in parallel, improving reliability of automated code workflows.