Don't call the agent. Give it something to react to. The supported pattern: define the trigger in AI Agent Studio (record trigger, scheduled trigger), then let your UI Action create or update the record the trigger watches. The UI Action calls a Script Include, the Script Include sets up the state, the agent fires on the trigger. Decoupled, testable, and it survives platform upgrades — unlike direct calls into agent internals. If you need parameters passed, put them on the triggering record (fields or a staging table), not in the call. "LLM instructions are not a data bus — if a value matters, a flow should carry it." Check the trigger types available in your release's AI Agent Studio docs before building — this area is moving fast and the exact API surface changes between releases. Build against the trigger contract, not the internals.
