“Lunch cost 35, and remind me at three to submit the documents.” After sending that message, you want to see an expense and a reminder. You also need a way to check the amount and time, then correct anything that is wrong.
In ISEA, the model interprets the request, tools carry it out, and cards show the results.
Turning language into an operation
The model needs to identify record type, amount, time, and action. Missing essentials may require clarification. Relative dates such as “tomorrow” must be interpreted against the request's time and time zone.
Tools check fields such as amounts and dates, verify account permissions, and then run the action. They return the result to the conversation, so a failed save can be reported even when the model understood the request correctly.
You can check both whether the AI understood you and whether the action finished.
Saving and presenting have separate jobs
Record tools perform writes and return results. Presentation tools organize information into useful interfaces. A polished card does not prove the database was updated; a successful write does not need a long paragraph repeating every field.
An expense card can show the amount and category. An import card can show a preview for you to approve. Details already visible on the card do not need to be repeated in a paragraph below it.
Native and interactive cards
Native cards use components the client already supports. The request declares available cards; the server checks the identifier and payload before the client renders them. These interactions can follow familiar app conventions.
Interactive cards allow more flexible layouts and actions. That flexibility remains subject to permitted actions and data interfaces. Interface code is not unrestricted database access.
Unsupported cards, invalid input, and rendering failures require compatibility feedback. A version mismatch should not leave the user guessing at an empty region.
Why imports need a confirmation interface
A bulk import may create many records at once. ISEA parses, validates, and prepares the complete result before asking the user to approve the write. A successful sample parse establishes only that some input is understandable; full parsing and persistence still have separate outcomes.
Confirmation connects the prepared result with the user's decision to add it. Cancellation leaves that import unwritten. Errors report the relevant stage. Repeated submissions reuse receipts within the supported identity scope.
Three questions the interface should answer
What stage is this at? What result already exists? Is an action required from me?
We aim to keep the page steady as text arrives, tools run, and results appear. Simple actions need only a brief status message. Cards are useful when you need to inspect details or make a decision.
Continue with safe record imports.