AI agents you can text: the chat thread is beating the app

Yesterday TechCrunch ran a roundup of AI agents you can text — assistants that don't live in an app you download, but in the same thread where your friends already are. You message them like a person, they remember the context, connect to your calendar and inbox, and actually do the thing: book the table, send the email, chase the follow-up three days later. The headline act, Instinct, just raised $1 billion at a $10 billion valuation. When investors put that kind of money on a texting interface, the app era is officially being challenged.

Why texting beats downloading another app

Think about the last five apps you installed. You opened each one maybe twice, then forgot the onboarding, the notifications and the whole thing. Now compare that with texting — the one interface everybody already knows, already has open, and already trusts with important conversations.

That's the whole pitch. Caddy sits in iMessage or RCS, reads the appointment hiding in an email, and drops it in your calendar. Comma is free and open source, lives in Signal or Telegram, and works across your files and browser. No new UI to learn is a feature, not a compromise — it's the same instinct behind why I give my apps away for free: remove every excuse between the person and the value.

The AI agents you can text right now

The roundup lists about twenty of them, and a few stand out:

• Instinct — the $10B one. It acts instead of just answering: trips, groceries, subscriptions, even its own dedicated email address so it can sign up for services on your behalf. It started placing calls in September. Still private beta.
• Poke — first AI agent approved on Apple Messages for Business, then acquired by Cognition in a deal valued in the low nine figures. From launch to exit in about four months.
• Fambot — a chief of staff for families. Every night it texts a summary of tomorrow: events, to-dos, what needs packing for school.
• Wajo's Fo — and this is the wild one — gets its own email address, phone number and payment card, so it can talk to businesses without borrowing your identity.
• Underdog — the privacy counter-argument: the whole model runs on your device, nothing leaves it.

An agent with its own credit card is a big deal

That Wajo detail sounds like a fun fact, but it's the whole industry's question in one line. If an agent can spend money, make calls and open accounts, who's accountable when it gets it wrong? Some builders lean into autonomy; others, like Maverick, ask permission before every action — which honestly feels like the right instinct. The article itself flags the privacy and security concerns, and I'd flag them louder. Trust isn't a feature you ship in v2; it's the product.

What I take from this as a builder

I build AI assistants myself — the four that live inside what I'm building taught me the same lesson this roundup confirms: the model is the easy part. Context, memory and knowing when not to act are what make an assistant feel like a colleague instead of a toy. Texting just happens to be the interface where all of that feels natural, because a chat thread already has memory, already has context, and already has a person on the other end who expects a reply.

My prediction: within a year, "download our app" will sound like "please visit our website" did in 2015 — technically an option, spiritually a defeat. The thread is where software goes to become invisible.

Would you text an AI to book your dentist appointment? Because plenty of people already are.