First agent mention (if any). Conventionally the routing target.
All entity mentions (#Entity / @{"type":"entity",…}).
All mentions, in order of appearance.
All skill mentions (/skill-name in the composer → @{"type":"skill",…} in message text).
Callers map these to ExecuteAgentParams.requestedSkillIDs; the server intersects them
with the agent's accepted skills AND the user's Run permission before any activation.
All user mentions.
Result of parsing all mentions out of a message.