One turn, start to finish, with nothing skipped over. The short version is that the agent can read a great deal and can write nothing — it can only ever hand you something to sign off, and the list of things it can hand you is three items long.
Plain language. No syntax to learn, and no need to know which table the answer lives in.
The model picks from a fixed catalog of tools — it cannot invent a new one, and none of the ones it can pick writes anything. Every call it makes carries your OAuth token, so the instance answers it exactly as it would answer you.
A row you aren't allowed to see comes back as a 403 — your instance's access controls refusing, not a rule of ours. We wrote no authorization logic; there is none of ours to audit.
Under every answer, while you work. Platform AI is metered in credits with no published price; this is a number in cents, from the same function the product uses to bill you nothing — you pay your model provider directly.
The card shows the exact payload — not a summary of it, not a description of what it will do. What you read is what gets written.
The body is editable in place. Discard and nothing was written — no draft queued anywhere, no record touched, nothing to clean up. This is the beat most products don't have.
The commit runs on your credentials through one named endpoint, and lands in your instance's own audit tables under your name. Not "sent by an integration". Sent by you, because it was.
The agent cannot invent a capability. This is the entire list it can choose from, and it is short enough to read in a minute — which is the point, because a security reviewer should not have to take a vendor's word for what an agent can do. However many proposals come back in an afternoon, each one stops at the same place: a card, and a person.
Anything you can do on your instance, it can draft — one record per card. Each of these returns a card and tells the model, in its own tool description, that it wrote nothing. The write happens later, if a human clicks, through the matching endpoint below — and not otherwise.
There are no background jobs, no scheduled runs, and no agent sitting on a queue overnight. Kaddiya works while a human is watching it work — which is the only honest way to sell a product whose safety property is that a person approves every write.
The technical consequence is the part we'd rather you check: unattended operation would require us to hold a durable token to your production instance, long after you've closed the tab. We don't want to hold one, so we didn't build the thing that needs it. Your session ends, the token goes with it.
A background agent is not on the roadmap as a feature. It would be a different product, with a different security story.