AI Glossary

Agent guardrails

Also called guardrails for agents

What does “Agent guardrails” mean?

Agent guardrails: Checks placed around an agent to keep it safe and on task: filters on inputs and outputs, limits on spending or steps, blocked actions, and validation of tool arguments. Guardrails reduce risk but cannot guarantee correct behavior.

Agent guardrails visual guide: Checks placed around an agent to keep it safe and on task: filters on inputs and outputs, limits on spending or steps, blocked actions, and validation of tool arguments. Guardrails reduce risk but cannot guarantee correct behavior.
Visual guide Open full size (new tab)

What is an example of “Agent guardrails”?

A guardrail might block any shell command that deletes files outside the project folder, regardless of what the model requests.

Keep learning

See also

Agent permissionsTermSandboxTermPrompt injectionTermGuardrailsTerm

Same topic

Autonomous agentTermHuman-in-the-loopTermVibe codingTermAgent loopTermAgent evalsTerm