You’re in the slow lane at the grocery store, cart half full, and you text a number you saved last week: “move my 4pm to Thursday, and remind me to buy coffee filters.” No app switch. No loading spinner. No login. Thirty seconds later a reply lands in the same thread where your sister sends you memes. That’s the whole interface.
I build bots for a living, and this is the part of the current agent wave I find most interesting. Not the demos with sixteen tool calls and a browser window. The boring channel. SMS and iMessage are the two surfaces almost every human on earth already checks, and a handful of teams have decided that’s where their agent should live.
Who’s actually showing up in the thread
TechCrunch put together a list of the most notable agents operating inside text messages, spanning general assistants to ones built specifically for families. A few names stand out for different reasons.
- Poke hit a real milestone in June 2026, when Apple approved it as the first AI agent on the Apple Messages for Business platform. That’s a distribution story as much as a product story.
- Ollie targets families and carries SOC 2 compliance, which tells you something about who they expect to be asking questions about data handling.
- Town positions itself as a professional assistant, handling tasks across a range of platforms.
- Genspark takes a different swing in 2026, offering unlimited access to multiple AI models for both chat and image generation.
- Cognigy sits on the enterprise side, leading voice with a native Voice Gateway, and shows up alongside Sierra, Decagon, and Salesforce Agentforce as a Tier-1 vendor-packaged option for enterprise customer support.
- CloudTalk is the odd one out, working the phone line rather than the text thread. Its AI VoiceAgents take inbound calls, check order and ticket status, book appointments, resolve FAQs, and route or transfer when needed.
Pricing is the gap here. The sources I’m working from don’t lay out numbers across the board, so I’m not going to pretend I know what any of these cost you at scale.
What the channel does to your architecture
Here’s where my actual opinion comes in. Shipping an agent into a messaging thread is not a UI port. It’s a different set of constraints, and most teams learn them the hard way.
You lose the screen
No buttons, no dropdowns, no modal asking “are you sure?” Every confirmation has to happen in prose. That means your agent needs an explicit confirmation step for anything destructive or expensive, written in a sentence a human reads in two seconds while walking. I’ve watched bots fire off a booking because the user replied “yeah ok” to something ambiguous. Design the confirm turn deliberately. Make the action restatable in under fifteen words.
Time stops being a variable you control
Web chat gives you a session. Text gives you a thread that lives forever and resumes whenever. A user can reply to your question six hours later from a different timezone. That pushes you toward durable state per conversation, idempotent task execution, and a clear policy on what happens when context goes stale. If your agent books the 4pm slot because it still thinks it’s Tuesday morning, that’s your bug, not the user’s.
Delivery is best effort
Messages get dropped, duplicated, and reordered. Carriers filter. If your task pipeline assumes exactly-once delivery of the user’s reply, you’ll ship double bookings. Treat inbound messages like events on a queue you don’t own.
Trust is the product
Ollie’s SOC 2 posture isn’t a marketing bullet, it’s a signal about where this category is heading. An agent in your messages sees your calendar, your family logistics, your purchases. The compliance work is table stakes once you’re handling that. And Apple’s approval of Poke for Messages for Business shows the platform gatekeepers are formalizing access rather than leaving it open. If you’re building on these surfaces, assume review processes get stricter, not looser.
What I’d build next
If I were starting a text-native agent this month, I’d pick one narrow job with a clear completion state. Rescheduling. Expense capture. Grocery lists. The enterprise players like Cognigy and the voice-first setups like CloudTalk succeeded by handling a bounded set of tasks well: order status, appointments, FAQ resolution, clean handoff to a human. That’s the pattern worth copying. Bounded scope, reliable handoff, no pretending the agent can do everything.
The general assistants are chasing a bigger prize, and some of them will get it. But the thread is an intimate place. One bad autonomous action costs you more trust than fifty good ones earn back. Build like the user is in a checkout line with one hand free, because they are.
🕒 Published: