A production deploy is releasing a code change to the live system that real users rely on. It is the step between code that has been written and tested and software that is actually running for customers.
It matters because mistakes at this stage have immediate effects. A faulty release can cause outages or data problems, so teams usually add checks such as tests, reviews and approvals before it happens. When an AI agent can start a deploy, human approval for that step is a common safeguard.
The process differs between organizations. Some deploy many times a day with automated checks and easy rollbacks, and others do so rarely with manual sign-off. Whichever approach is used, the ability to deploy is a powerful permission that should be limited to the identities that need it.
To manage the risk, restrict who and what can deploy, require review and automated tests first, make rollbacks quick and log each deployment with the identity that started it.