What are AI guardrails used for?#
Guardrails act as a real-time safety net, actively guiding the AI's responses to keep them aligned with company policies, brand voice, and ethical standards, thereby increasing the system's reliability and trustworthiness.
How do AI guardrails work?#
Guardrails function by implementing a series of checks and rules that the AI must follow. These can be simple keyword filters that block certain topics or language, or more complex instructions that define the AI's behavior, such as "Do not provide financial advice" or "Only answer questions related to our product catalog."
These rules are applied in real-time as the AI operates. The system might check a user's request before the AI processes it, monitor the AI's response as it's being generated, or review the final output before it is delivered to the user to ensure it complies with all predefined policies.
Guardrails for coding agents: hooks#
In agentic coding, the instructions a team gives an agent (an AGENTS.md or CLAUDE.md file,
rules, skills, prompts) are suggestions. The model follows them most of the time, which is fine
for style and not fine for a rule that protects secrets, production configuration or the release
pipeline. As Ran Isenberg puts it in Agentic Coding Hooks: Deterministic AI Guardrails,
"most of the time is a probability, not a guarantee."
Hooks are the deterministic layer. They are ordinary scripts that the agent's runtime runs at
fixed points in its lifecycle, such as before a tool call (PreToolUse in Claude Code) or after
one succeeds (PostToolUse), and "the model doesn't get to decide whether your code runs. The
runtime does." A pre-tool hook can inspect the command an agent is about to run and block it,
for example a recursive delete or a read of an .env file, with the reason passed back to the
model. A post-tool hook is commonly used to run the formatter and linter after every edit.
Cursor, OpenAI Codex, Gemini CLI and GitHub Copilot CLI have their own hook systems.
Guardrails are one half of a delegation decision. They enforce what must not happen; the delegation boundary states what may happen without asking, and a contraction trigger is the written condition that narrows it again — a model change on a consequential path, a rising exception rate, a near miss. Widening it is automation follows trust: a check becomes automatic once the signal behind it has proved reliable, not before.
Hooks work best reserved for the few outcomes that must never happen, with everything else left to instructions. Because a hook is code that runs automatically, a project's hook configuration deserves the same review in a pull request as its CI configuration.
How are AI guardrails different from fine-tuning?#
Guardrails are a real-time safety system applied to an AI's outputs, distinct from the process of training the model itself. While fine-tuning alters an AI's core knowledge by retraining it on specific data (like teaching it your company's product manual), guardrails act as a separate layer of rules that filter or steer its behavior on the fly without changing its fundamental training.