What is not live yet

Designed, visible in the product, and not switched on. Each one is marked where you would meet it.

The list the product itself reads

This is not a documentation summary. It is the actual list the product consults when it has to tell somebody that part of what they asked for will not happen — the same words appear in the builder, in the note after an agent is written, and on the cards of the diagram. A test holds this page to that list, so switching one on here without switching it on there, or rewording one and not the other, fails the suite.

  • checkout, the product says: “completing the payment itself, so it fills the basket and stops at your approval”.
  • schedule, the product says: “starting a run on that schedule, which is saved and has nothing to run it”.
  • site-signin, the product says: “signing in to those sites, which is built and has not yet worked on a real one”.
  • history-study, the product says: “reading the last ninety days and suggesting from it, which is built and has not yet run on a real mailbox”.
  • business-signup, the product says: “connecting a business's own mailboxes, which is built and has not yet read a real one”.
  • backstop, the product says: “the backstop that holds a commitment nobody approved, which is built and tested against a fixture and has not yet stopped one on a real shop”.
Also not built
  • Starting a run on a schedule. A cadence is part of an agent's definition, the diagram draws it, and the Scheduled tasks page lists what is queued — but nothing turns a cadence into a run. A task sitting there at its due time will stay there. Pausing and deleting it work now, because those are what you need the moment a clock is switched on.
  • Completing a purchase, even with a card already saved at the shop. An agent can fill the basket and stop at your approval. How paying would work when it is built is written down in docs/wallet-decision.md, and the answer is a card with your approved amount as its ceiling rather than one of yours.
  • Sending email out. The path is written and needs a provider key and a verified sending domain.
  • Deleting your account and its data on request. The button records a dated request and says, in those words, that nothing has been removed.
  • A feedback inbox. The Feedback page offers two routes that reach a person instead of a form that reaches nobody.
  • Reading your mailbox. Setting a business up has seven screens and the study that fills screen four runs in the engine's worker, which this installation does not have. Everything those screens show runs on a made-up company instead, and says so at the top of every screen that shows a figure.
  • The limits on what an agent may read. Mailboxes, senders and labels you rule out on screen six are stored and nothing consults them. The screen says so before you fill it in.
  • The pause switch on an agent. Pausing is stored and the runtime does not read it, so an agent paused on its page can still be started. The agent's page says so while it is paused.
  • What an agent knows. The page lists every fact and lets you take any of them away; no run consults the list yet.
Deliberately not attempted
  • Anything off the web: files on your computer, a terminal, a native app, a coding session on your laptop.
  • Solving MFA prompts or CAPTCHAs on your behalf.
  • Taking over a live browser video stream. Runs show stored screenshots, which are labelled as such.
  • Arbitrary code, loops or parallel branches. The graph is a bounded set of steps with one route taken at a time, because there is no sandbox to run arbitrary code in.
Recently switched on

Two things left this page rather than joining it, which is the direction it is meant to move in. Texting works and has been run end to end on a real phone. Signing in to an ordinary website works: a password is kept encrypted and typed only on its own site, a texted code is asked for when the site sends one, and the one way in that is still yours to do is Continue with Google or Apple, a passkey or a captcha.

The entry for signing in is worth naming, because it is the failure this page exists to prevent. The product went on telling people their accounts could not be connected for a while after they could, because the sentence lived in one place and the feature lived in another. It now lives in one place and both read it.

Why it is shown at all

A feature you can see and cannot use is better than one that is hidden, as long as it says so. Hiding it means the first time you learn the product will not do the thing is after you have built something that depends on it.