Instinct's personal agent grows 10% daily, limited only by compute
Instinct bets that a personal agent reached through text, voice and email, with no new app, can become the default way people handle errands and money. Users hand over more sensitive data as trust builds, and that data keeps them, while transactions such as travel give a path to revenue. Compute supply, not demand, is the limit on how fast it grows.
Takeaways
- Instinct is a personal agent without an app. Users reach it by text, call or email, and it has its own phone and computer. It can call about deadlines and act across the internet. More than 50% of traffic skips iMessage, and some users send over 90% voice.
- Invites drive growth of about 10% daily. The company is about a year old and invite-only, with five invites per user and no marketing spend. Daily growth rose from 1% and 2% to about 10% after a couple thousand users.
- Trust builds over weeks and lifts retention. Three weeks in, about 40% of users have shared a personal credit card. Users who connect at least one sensitive item retain at 80%. Shinn treats the ability to take access back as a core principle.
- Travel anchors transaction volume. Travel is 50% of transaction volume, and the platform is approaching over a billion dollars a year. One voice request can book a flight, hotel and rides, which Shinn says could reshape travel agencies and restaurant reservations.
- Compute, not demand, is the binding constraint. Demand is doubling roughly weekly while compute has a lead time of several months. Buying ahead can cost three or four times more if the forecast is wrong. Shinn spends about 40% of his time on it.
- Revenue will be a take rate, not ads. Shinn rejects steering users for advertisers. He expects a blanket transaction take rate, as with payment rails, Shopify at 2.5 to 3 percent, or Apple's 30%. He has not found where Instinct lands.
- Incumbents will be changed, not disrupted. Shinn says most digital services will be changed rather than disrupted, so Instinct should work with incumbents. He advises businesses to run scaled experiments, for example enabling a change for 1% of users.
Safeguards added after an early version
- An early version lacked firewalls and active monitors. Shinn says the team built systematic fixes rather than patches.
- Firewalls screen incoming content before the agent acts on it.
- A monitoring system decoupled from the agent can pause or block actions.
- A validator catches invented proper nouns before a tool call.