Glossary
Guardrail checkpoints (input, action, output)
The three places in the agent loop where guardrails run: on the input (what the person asked and what retrieval brought in), on the action (the tool call, its arguments and its result) and on the output (the answer and what it renders).
NeMo Guardrails calls them input, retrieval, execution and output rails; the OpenAI Agents SDK has input, output and tool guardrails whose tripwire halts the agent. They catch the obvious and miss the crafted, which is why none of them should be the only defense.