Guardrails
Guardrail
Checks that stop or constrain an action before it happens.
Guardrails are the rules that constrain what an agent is allowed to do: a path it may not write to, a command it may not run, a class of request it must refuse, a number it may not exceed. They exist because the model cannot be trusted to police itself; it will attempt the blocked thing if nothing stops it.
Their placement matters. A guardrail on the output catches bad answers after the model has already produced them, and the user has usually already read them. A guardrail on the tool call catches the action before it happens. A guardrail on the input catches a prompt injection before it reaches the model. The strongest systems check all three.
Guardrails also need to be narrow. A rule that is too broad turns the agent useless, and a rule that is too specific misses the case it was written for. This is the same lesson as ordinary security: allow by default, deny by exception, and log the exceptions.
Related terms
Harness
Everything around the model that turns it into an agent: tools, system prompt, permissions, hooks.
Tool Call
The model's request naming a tool and its arguments. Structured output, not a side effect.
Prompt Injection
Text that tries to make the model ignore its instructions and follow the attacker's instead.
Human in the Loop
A person who reviews, approves, or corrects the agent's work at key points.
Guardrails
Checks that stop or constrain an action before it happens.
Want this working inside your team?
It starts with one conversation about your workflow.