Autopilot watches your tasks and does the work you allow it to. It starts in shadow mode, where it only shows you what it would have done, and it asks before it finishes anything.
The rule builder
Signal
How much a computer can do
is mostly or fully
Signal
Needs something only you know
is false
Logic
All of (2)
one rule
ActionL3
Draft it
with your key, tell you in the digest
Scope
Crumbs, in all tasks
when a task is planned
What this produces
1 rule
When a task is planned, for every crumb a computer can mostly do that does not need you, draft it and tell you in the digest.
Guardrails
Up to 20 actions a day. Never past "ask me first".
Signals
Facts about a step. How long it should take, what kind of work it is, whether it needs something only you know.
All of
Everything wired into this has to be true. Swap it for Any of and the rule splits into one rule per branch.
The action
The rung of the ladder Autopilot may climb. L3 drafts something for you to finish and never marks anything done.
Scope
Where the rule applies and when it is allowed to look. Crumbs in every task, checked each time a task is planned.
What this produces
The panel reads back the rule in a sentence, counts how many rules your drawing makes, and shows the limits on top.
The short version
You draw a rule. Autopilot reads it back to you in a sentence. For three days it runs in the background and writes down what it would have done, without doing any of it. You read that list, and then you decide whether to switch it on.
Even switched on, it stops and asks you before anything is marked done. That is not a limitation we plan to remove quietly. It is the point.
Autopilot works on crumbs, not on tasks
When a task is broken into steps, each step gets classified: what kind of work it is, roughly how long it should take, how much of it a computer could do, and whether it needs something only you know. Those are the facts your rules read.
That is why Autopilot can be useful without being reckless. It is not deciding what to do with your goal. It is deciding what to do with one small, already described step.
A rule is four pieces
The panel on the right turns your drawing into plain sentences as you go. If it does not say what you meant, the rule does not do what you meant.
Shadow first, always
"Save and turn on" does not turn it on. It saves the rule set in shadowfor three days. Autopilot evaluates everything, writes a line in Activity for each decision, and does nothing. The lines read "Would have drafted this".
After three days you have a real list of what this rule would have done to your real tasks. Switching it on for real is a separate, deliberate step.
Nothing finishes without you
The highest rung Autopilot reaches today is "Complete, then ask me". It does the work, attaches the result, and moves the step to awaiting approval. One tap in the app, the email or on your phone approves it. Until you tap, nothing is done.
There is a rung above that on the ladder, where a scoring step decides for you. It is not available: the scoring step is not built, and while we are launching, every result waits for a person regardless of settings. See guardrails for exactly what that means.
Six rule sets to start from
You do not have to draw anything from scratch. The gallery ships six, including "Research every new crumb", "Draft what a computer can do" and "Unstick skipped crumbs". Load one, read what it says, change the parts you disagree with.
What it costs to run
Autopilot uses your own AI key, the same one decomposition uses. Every run records how many tokens it used and an estimated cost, and Test Run shows you that estimate before anything runs at all.
Keep reading