Bloquer ou rappeler : les deux familles d'outils d'un agent
Two ways to equip an agent, and they do not age the same way. A **blocking** tool makes up for judgement you do not trust yet: it forbids editing a given file or running a given command. A **reminding** tool nudges at the right moment, with a re-injected instruction or a warning, and leaves the call to the agent. The first turns into friction as the agent improves, like the restrictions placed on a junior who has since become senior. The second keeps paying off, because a senior still slips now and then, and also knows when a deviation is deliberate. The distinction comes from Daisy Hollman, who works on Claude Code at Anthropic, in a video captured on 10 September 2026. In Claude Code both kinds run through the same mechanism, hooks: a `PreToolUse` hook that exits with code 2 blocks the tool call, while a hook that returns `additionalContext` slips an instruction into Claude's context without stopping anything.