Why your AI coding agent keeps ignoring your rules—and how to fix

The core argument is that engineering teams over-rely on rules for AI coding agents, but rules only guide probabilistic LLMs. Hooks—deterministic scripts triggered by agent events—provide actual guarantees. The speaker demonstrates that bloating rules degrades agent performance, citing a study where self-evolved rules made agents worse.

The practical taxonomy is clear: pre-tool hooks block actions (security gates), post-tool hooks log activity (observability), and stop hooks enforce completion criteria (e.g., mandatory test passes). The most actionable insight is the 'rule audit' framework: classify each rule as encoding judgment (keep as rule) or naming a process/extracting a guarantee (convert to hook).

The speaker walks through concrete implementations: blocking .env file reads, forcing file coupling checks before edits, and running tests on conversation end. These are not speculative—they use exit codes (0=pass, 2=block) that agents interpret as feedback, not crashes.

The tradeoff is clear: hooks add setup complexity but remove failure modes from probabilistic agent behavior. For teams scaling agent usage, the recommendation is to audit rules now, extract process guarantees into hooks, and accept that this requires initial engineering investment.