It’s 7:12 a.m. You’re standing in the kitchen with one hand on a coffee mug and the other thumbing out a message. Not to a person. To something that will book the thing, remember the other thing, and nudge you at 4 p.m. because you asked it to. No app icon tapped. No login screen. No onboarding carousel. Just a gray bubble, a blue bubble, and a response.
That’s the whole pitch for texting agents, and as someone who spends most days wiring up bots, I find it quietly radical. Not because the models got smarter, but because the delivery surface got dumber in exactly the right way.
Messaging is the lowest-friction runtime we have
Every bot I’ve shipped into a chat thread has outperformed its web equivalent on one metric that matters more than any benchmark: whether people actually come back. The reason isn’t mysterious. Your messages app is already open. It’s already pinned to your dock and your lock screen. It already has notification permissions you granted years ago and never revoked. A bot that lives there inherits all of that for free.
Compare that to shipping a mobile app. You need a store listing, a review cycle, an install, a signup, a push-permission prompt, and then a reason for someone to find your icon among eighty others. Four of those five steps are places users quietly disappear. In a text thread, the funnel is a single step: send a message.
What the Apple approval actually signals
Poke launched in March 2026 as a general-purpose, text-based assistant. In June 2026, it became the first AI agent approved on Apple’s Messages for Business platform. If you build bots, that second sentence is the interesting one.
Messages for Business isn’t an open pipe. It’s a gated channel with a review process, and historically it’s been where airlines and banks go to handle support tickets. An AI agent clearing that bar means Apple is willing to treat a general-purpose assistant as a legitimate conversational endpoint rather than a novelty. For builders, that reframes the integration question. The hard part of a texting agent was never the model call. It was distribution and the carrier and platform plumbing underneath it: SMS versus RCS versus iMessage, delivery receipts, opt-outs, rate limits, and who gets blamed when a message lands twice.
A sanctioned path through Apple’s platform is the kind of unglamorous infrastructure win that makes a category viable. It also sets a precedent that review processes will exist, which means the next wave of texting agents will be designed around approval constraints from day one.
Pricing tells you more about the architecture than the marketing does
Ollie, a family-focused assistant, offers a free tier plus paid plans starting at $25 per month for 150 messages, with a $100-per-month plan covering up to 1,000 messages.
Sit with those numbers for a second, because they’re a design document in disguise. Pricing per message rather than per seat or per month of unlimited access tells you the unit of cost is the turn, not the user. Each inbound text likely triggers tool calls, context retrieval, and possibly a human-grade action with real consequences. At roughly 17 cents a message on the entry plan and 10 cents on the larger one, you can see the volume discount doing its work, and you can infer that the per-turn cost is meaningful enough that nobody wants to eat it on an unlimited plan.
If you’re building in this space, that’s a useful calibration. Metered conversation pricing pushes you toward agents that do a lot per turn instead of agents that chat. A bot that resolves a request in two messages is worth more to the user than one that asks four clarifying questions, and it costs you less. The incentives line up, which is rare.
Trust is the real moat
Ollie is also one of the first mainstream, family-focused AI assistants to achieve SOC 2 compliance, a widely used security standard. That sounds like a procurement checkbox until you remember what a family assistant touches. Kids’ schedules. Doctor appointments. Addresses. The contents of your actual conversations.
An agent in your text messages sees your most intimate channel by default. It’s the same place your partner messages you and your bank sends codes. Any product that wants to live there permanently needs an answer to “where does this data go” that’s stronger than a privacy policy paragraph. Pursuing a recognized audit standard is a solid signal about where a team thinks the competition will actually be decided, and I suspect they’re right. Model quality is converging. Handling of sensitive context is not.
What I’d build next
If I were starting a texting agent this week, I’d optimize for three things in this order: resolve requests in the fewest turns possible, make every action reversible or confirmable before it fires, and treat the message log as sensitive storage from the first commit rather than a retrofit.
The gray bubble is a constrained interface. No buttons, no dashboards, no place to hide a complicated flow. That constraint is the feature. It forces you to build an agent that’s genuinely good at understanding intent, because there’s nothing else to lean on.
🕒 Published: