A write boundary is the line between what an agent may read or prepare and what only a person may complete. On one side everything is reversible. On the other, something is spent, sent, or published under your name.
Why it matters
Agents have become genuinely good at the preparation. They can assemble the order, draft the brief, pick the option that fits, and fill in the form correctly on the first try.
What changes the stakes is the last click. A purchase, a publish, a send: those carry a cost and a reputation, and an agent free to take them on its own is a new kind of exposure for the brand it works for.
Drawing that line in policy is not the same as drawing it in code. A rule an agent is asked to respect is a rule it can be argued past. A boundary it structurally cannot cross holds under any prompt, including a hostile one.
How Jinn treats it
On the Jinn gateway an agent can queue a purchase and never complete one. The single tool that changes state returns a checkout link or an intake form and hands control back to a person, and it is built so it cannot complete a payment on its own.
Everything else reads. That split is the point: an agent does the work of preparing, and the decision that costs you money stays where it belongs, with you.
An agent can queue a purchase but never complete one; a human approves every checkout. See it
Related terms
Agents prepare. People decide.
The last click stays yours, not by policy but by how the thing is built. See how agents connect to a brand.