Guardrails

Guardrails are the constraints around an AI system that stop it doing things it shouldn't — restricting which tools it can call, validating its output against a schema, filtering topics, and capping how many steps or how much money it can spend.

The most effective guardrails are the least sophisticated. Give an agent read-only database access and it cannot delete anything, regardless of what it decides. Validate output against a schema and malformed responses are rejected before they reach a downstream system. Cap the step count and a confused agent stops instead of looping.

Prompt-level guardrails — "never discuss competitors" — are the weakest kind, because they're instructions to a system that probabilistically follows instructions. Useful as a layer, never as the only one.

Why it matters

Guardrails are what make the difference between an AI feature that's safe to expose to customers and one that's a liability. They're also usually the part quietly missing from a cheap quote.

Related terms

AI AgentAn AI agent is a program that uses a language model to decide what steps to take toward a goal, then carries them out by calling real tools — sending …Human in the LoopHuman in the loop means an automated process pauses for a person to approve, correct or reject before it continues — typically before anything irrever…HallucinationA hallucination is output that is fluent, confident and false — an invented citation, a policy that doesn't exist, a fabricated order number. It's a d…

More in AI Automation & Agents

AI AgentWorkflow AutomationLarge Language Model (LLM)RAG (Retrieval-Augmented Generation)EmbeddingVector DatabaseHallucinationHuman in the LoopTool Use (Function Calling)Model Context Protocol (MCP)PromptTokenEvalsIdempotencyWebhookAPI IntegrationAI Voice Agent

Need this built rather than explained?

We publish every price we charge, and you get a quote in writing before anything starts.

See every price
All terms