How Twinly executes

A task drops through several layers, fastest and most certain first, and only reaches the model if nothing simpler can do the job:

1. Recipe: a built-in operation for common web actions, with no AI call at all.
2. Learned pathway: steps the twin learned from a past run and now replays, with no AI call, and self-healing.
3. Full AI run: the twin reasons through a genuinely new task, and its success teaches the layers above.

Alongside them sit the right way in (how each step reaches an app) and verification (how the twin confirms an action happened). Underneath all of it is memory, fed into every run.

A recipe is a hand-built operation for web actions whose path never changes: a site search through its standard address (Google, Reddit, YouTube, Amazon, eBay, Wikipedia, GitHub, Stack Overflow, Maps) or a direct open (Gmail, Calendar, Drive, LinkedIn, and more). There is nothing to infer, so the twin goes straight there.

The matcher only fires when it is confident, so a nuanced request is never hijacked by a recipe that half-fits. And it composes: a simple search finishes right there, but if your request has a follow-on (text it to me, save the top five to a sheet), the recipe opens the right page and hands off to the layers below.

After the twin completes a task by driving your apps, it records the steps as a reusable pathway. Anything that came from your wording becomes a fill-in-the-blank, anything secret is dropped, and commit buttons (Post, Send, Buy) are kept exactly as written so the same rules apply on replay. The next time a matching task arrives, the twin replays the saved steps instead of working it out again.

• A new pathway is still learning until a clean replay proves it trusted.
• A single failed replay drops it back to learning.
• If anything drifts mid-replay, it stops and falls through to a full AI run, which recovers the task and records a corrected version.

Every step reaches an app through the best way in it offers:

• An app's own interface first: Calendar, Reminders and Contacts through the system, PDFs edited directly, apps that can be scripted through their own scripting, and Connectors.
• On websites, the page itself: the twin lists the page's controls, fills and clicks them by the words they show, and reads each form back after a change. See Websites and apps.
• In other apps, the screen: a screenshot after every step plus the window's own controls and text, clicked and typed with your real pointer and keyboard.

After a committing action, the twin checks instead of taking the tool's word: a created calendar event is confirmed in your calendar, a Gmail send in your Sent box, a written file on disk with the right content, a text's delivery in Messages. When a check proves something did not happen, the twin tells you (couldn't confirm it went through, check before relying on it). When an action simply cannot be checked, it reports exactly what the tool said.

When no recipe and no trusted pathway matches, the full agent runs on Twinly's engine (see Powering your twin). Its successful runs feed back up: a new task becomes a learned pathway, so the next similar one is faster and more certain.