Most AI can say “I’ll take care of that.” The interesting question is what exists in the world five seconds later.
Here, the answer is a Plan — and the fact that there’s an object rather than a sentence is most of what separates this from a chat window.
A Plan is a contract, not a to-do
When you commit to something, or Aura proposes something worth pursuing, it becomes a Plan with a clear objective, a next action, a time that action happens, and a way to verify it actually did.
That last part is the one that matters. Delivery is not completion. A reminder that fired is a reminder that fired; the Plan closes when its condition is actually satisfied. The difference sounds pedantic until you’ve used software that congratulates itself for sending you a notification you ignored.
Plans show their state plainly: on track, needs you, paused, done. You can see them, and you can see what’s scheduled behind them.
”I’ll…” is a promise with machinery behind it
The rule the system holds itself to: if Aura says I’ll do that, there is scheduled work behind the sentence, or the sentence doesn’t get said.
This is worth knowing because it changes how much you should trust the phrase. In most products “I’ll remind you” is a hopeful noise. Here it’s a claim you can check — the scheduled action exists, or Aura tells you it couldn’t create one.
Approvals: the edge, and why it stays
You appear at the irreversible edge. Send, submit, buy, post, call a real person. Everything upstream — finding, researching, drafting, sequencing — happens without you, and what arrives is finished work with one thing to approve.
People sometimes read that as a limitation the mature product will remove. It isn’t, and I want to be direct about why: an agent that can commit your money and your reputation without asking is not a more advanced agent. It’s the same agent with the brakes off. Symbiosis makes the same argument about commerce, and it holds here.
The engineering problem was never how do we stop asking you. It was how to make the moment worth the interruption — arrive with the work genuinely done, so the decision takes five seconds and is actually informed. A yes/no on a finished draft is a different act from a to-do list handed back.
Approvals wait in one place, so the answer to “what needs me?” is a list rather than a hunt.
When something fails
The part I’d point at if you only read one section.
Things go wrong: a connector expires, a service is down, the thing you were waiting on never came. What Aura does then is say so. A blocked Plan says it’s blocked and why. A delivery that couldn’t be verified says it’s delivered but unconfirmed, rather than claiming success.
It never fakes a finish. That is the single behaviour I’d want you to test in your first month, because it’s the one that decides whether any of the rest is worth trusting — a system that quietly reports success it didn’t achieve is worse than one that does nothing, since you stop checking.
Living with it
Let the first few Plans run without micromanaging. The instinct is to check on them. Checking teaches you nothing that the outcome won’t.
Kill the ones you don’t mean. A Plan you’re no longer interested in should be closed rather than left to nag — and an honest “not worth acting on, revisit in March” is a legitimate ending the system supports.
Watch the approval queue rather than the chat. Once a few Plans are running, “what needs me” is the real interface. The conversation is where things start; the queue is where your day gets decided.
Notice the one-suggestion rule. Your morning brief carries exactly one suggestion, never a list. That’s a deliberate constraint on the system’s right to spend your attention, and it’s the thing that keeps a proactive product from becoming a noisy one.
Next, and last: Insights — letting it work while you don’t.