In practice
A media stream forks or redirects a call's audio to a URL you control. Bidirectional streams let the application play audio back into the call, which is how many voice AI agents speak with callers.
Streams typically carry compressed telephony audio in small frames. The application has to handle speech recognition, turn-taking, interruptions and latency itself, since every delay is audible to the caller.
Media streams and SIP trunking are the two common ways to connect a voice AI agent to phone numbers. Streams are easier to start with; SIP fits teams that already run their own voice infrastructure.