Local code and on-device classifiers decide when something’s worth acting on. Agents see one message, never your whole day.
“When I promise something in Slack, add it to my todos.”
Plain code reads your apps. An offline model answers: does this match?
One message, to the agent you chose. Nothing else.
Follow-ups drafted, promises tracked, asks answered.
A local model spots what the task didn’t ask for.
Keys, tokens and personal info never reach the model.
Your agent still gets the real value.