Glossary
Approval gate (action checkpoint)
The check between an agent's decision and the system it touches: the policy says which tools and arguments are allowed, and a person approves the actions that are sensitive or irreversible.
OpenAI's guide rates each tool's risk and makes high-risk, irreversible actions a trigger for human oversight; Anthropic's guide has agents pause at checkpoints for human feedback. Auto-approve and unattended modes remove the gate on purpose, and most of the damage in the incident library went through it.