hooks

A hook has three possible answers to an agent's action: allow it, ask for an approval, or block it. Hugo Lassiège, documenting his solo software factory on 28 July 2026, makes that the whole argument for the mechanism: « Ce qui compte doit être exécutable. Une consigne est suivie "la plupart du temps"… Un hook ou un test est suivi tout le temps. » In his grid, hooks answer the third question he puts to any piece of tooling, what stops the agent when it gets something wrong, next to tests and CI gates. The harness fires them.

Louis Claxton, of Anthropic's Applied AI team, splits governance the same way in his AI-native SDLC playbook of 21 August 2026: the skill is an advisory control, the hook is the deterministic layer that blocks or demands an approval. That division is what turns Claude Skills into something closer to policy than to suggestion. The attachment points are the PreToolUse and PostToolUse events, and the Claude Code curriculum published in February 2026 counts more than eleven such events, filed under advanced customization next to subagents and custom slash commands.

Addy Osmani lists hooks among the seven primitives of the harness on 19 April 2026, with their design rule attached: "success is silent, failures are verbose". He also records the Anthropic line that harnesses do not shrink, they move, as useful primitives become standard and get post-trained into the next model. Whether a layer whose value is being non-negotiable can migrate that way is left unanswered.