Agent hooks

Agent hooks inject deterministic code execution into probabilistic AI agent workflows to enforce security, automate quality checks, and manage context.

Relying on system prompts to keep AI agents in check is a losing battle. Agent hooks solve this by introducing deterministic, code-driven guardrails directly into the agentic loop (such as in GitHub Copilot, Claude Code, and VS Code). By registering custom scripts to trigger at precise lifecycle events (like PreToolUse or PostToolUse), developers can programmatically block destructive shell commands, run linters immediately after code generation, or inject real-time context. It moves critical operational policies out of the model's unpredictable memory and into guaranteed, executable code.

https://code.visualstudio.com/docs/copilot/customization/hooks