Think about the last time a restaurant handed you a QR code menu. You didn’t want a new interface. You wanted the food. Most software we build has the same problem: we hand people a menu when all they want is the meal. Text-message agents flip that. There’s no onboarding flow, no empty state, no settings screen with a toggle nobody touches. There’s a thread, a contact, and a blinking cursor people already know how to use.
That’s why the current crop of textable agents interests me more as a builder than as a user. Poke, Ollie, Orbits, and Instinct are all chasing the same insight from different directions, and the differences in how they package it tell you a lot about where this kind of product is headed.
Distribution is the real moat
Poke’s milestone is the one I keep coming back to. In June 2026, Apple approved Poke as the first AI agent on the Messages for Business platform. If you’ve ever tried to get software in front of users through a channel you don’t own, you understand what that means. Being first through a door that Apple controls is worth more than almost any model improvement you could ship in the same quarter. Poke was later acquired by Cognition, which reads less like an exit for a chat app and more like someone buying a distribution channel with a general-purpose assistant attached.
For anyone building in this area, that’s the lesson. Your agent’s quality matters, but the surface it lives on determines whether anyone meets it. Text messaging is the rare channel where every phone on earth is already a client.
Trust shows up as paperwork
Ollie takes a different route. It’s a family-focused assistant, and it’s SOC 2 compliant. That second detail sounds boring until you consider what a family assistant actually handles: schedules, locations, kids’ activities, who’s picking up whom. The moment your agent touches that kind of data, compliance stops being a sales checkbox and becomes part of the product.
I’ve watched plenty of side-project bots die at exactly this point. The prototype works beautifully, then someone asks where the message logs live and for how long, and there’s no answer. Ollie offers a free tier with paid plans starting at $25 per month for 150 messages. Those are not casual-hobby numbers. That’s pricing that assumes real infrastructure underneath, including the audit trail.
Message counts tell you about the architecture
Compare that to Orbits, which runs a free beta and a Pro plan at $8.33 per month for unlimited tasks. Metered messages versus unlimited tasks is not a marketing choice. It’s an architecture choice leaking into the price sheet.
When you charge per message, you’re signaling that each exchange carries meaningful cost: inference, retrieval, possibly human review. When you charge a flat rate for unlimited tasks, you’re betting your per-task cost is low and predictable, or that most users will never approach your ceiling. Both bets can work. Both also shape what you build. Metered pricing pushes you toward fewer, denser replies. Flat pricing pushes you toward cheap routing, caching, and aggressive early exits from expensive paths.
If you’re designing your own texting agent, pick the pricing model early, because it will quietly dictate your prompt design, your model tiering, and how chatty your bot is allowed to be.
Money is arriving before the habits do
Then there’s Instinct, which raised $1 billion in September 2026 at a $10 billion valuation. Those are numbers attached to a behavior most people have not adopted yet. Investors are pricing in the belief that texting an agent becomes a default, not a novelty.
I’m sympathetic to that belief, with a caveat. SMS and iMessage threads are the most intimate software surface people have. A pushy notification in a Slack channel is annoying. The same message in the thread next to your mother is a violation. Agents that live here inherit an enormous amount of trust and have very little room to spend it.
What I’d build differently
A few things I’d keep in mind after looking at how these four are positioned:
- Design for interruption, not sessions. Nobody opens a text thread for a twenty-minute working session. Every reply needs to make sense in isolation.
- Treat latency as a feature. A thread tolerates a delayed answer far better than a web app does. Use that slack for better retrieval instead of racing to a worse reply.
- Keep memory explicit. Users should be able to ask what you remember and tell you to forget it, in plain text, without a dashboard.
- Own your compliance story before you need it. Ollie’s SOC 2 position is a competitive asset, not overhead.
The interesting race here isn’t over who has the smartest model. It’s over who earns a permanent spot in a thread people already read a hundred times a day. That’s a much harder thing to win, and a much harder thing to lose once you have it.
🕒 Published: