Remember when the big consumer AI bet was the chat window? A tab in your browser, a blinking cursor, a blank box waiting for you to think of something clever. That interface defined the first few years of the assistant era, and it quietly set the ceiling on what assistants could do: they could only help you when you remembered to go visit them.
The current crop of texting agents inverts that. Poke, Ollie, Orbits, and Instinct all live where you already are, in the thread next to your sister and your dentist’s appointment reminder. From an architecture standpoint, that’s not a cosmetic port. It’s a different machine underneath.
The channel is the constraint
SMS and iMessage impose a design discipline that web chat never did. There’s no sidebar, no file upload affordance, no settings panel, no streaming token animation to paper over latency. You get a few hundred characters and a send button. Everything an agent needs to communicate about its own state—what it’s working on, what it’s confused about, what it’s about to do on your behalf—has to be expressed in the same channel as the answer itself.
That forces a decision most agent frameworks dodge. In a web app, you can show a tool-call trace, a progress spinner, a confirmation modal. In a text thread, a confirmation request is just another message, indistinguishable in form from small talk. The agent has to maintain its own notion of pending state across an unbounded, asynchronous conversation where the user might reply in four seconds or four days, and might reply to something three turns back without saying so.
Anyone who has built a dialogue state tracker knows how badly that breaks naive turn-based loops. The useful agents here are necessarily doing something closer to event-driven orchestration than request-response inference.
Apple’s approval is the architectural signal
Poke became the first AI agent approved on Apple’s Messages for Business platform in June 2026, and its parent, The Interaction Company of California, was later acquired by Cognition. The acquisition is the headline, but the approval is the more interesting technical fact.
Messages for Business is a governed channel. Getting an autonomous, general-purpose agent onto it means satisfying a platform owner that the system’s outputs are bounded enough to carry a brand-adjacent identity inside Apple’s own app. That’s a constraint on the model layer, not just the product layer: output filtering, identity handling, and some defensible story about what the agent will refuse to do. Being first through that door is a moat built out of compliance engineering, which is an unglamorous but durable kind.
Pricing tells you where the compute goes
Compare the published plans and you can reverse-engineer the cost models.
- Ollie is family-focused, holds SOC 2 compliance, and offers a free tier with paid plans starting at $25 per month for 150 messages. That’s roughly 17 cents per message, metered by message.
- Orbits runs a free beta plus a Pro subscription at $8.33 per month for unlimited tasks.
Metering by message, as Ollie does, implies that each inbound text can trigger an expensive, open-ended chain of work. Offering unlimited tasks, as Orbits does, implies either tighter per-task bounds, aggressive caching and routing to smaller models, or a tolerance for subsidy during beta. Neither is wrong. They’re bets about variance. A per-message price protects against the user whose single text spawns a fifty-step plan; an unlimited price bets that most tasks are short and predictable.
Ollie’s SOC 2 posture matters for a related reason. A family assistant sitting in a text thread sees school schedules, addresses, health logistics. The moment an agent has persistent memory across months of messages, the memory store becomes the sensitive asset, not the model. Compliance certification is a signal about where that data lives and who can touch it.
Capital is flowing to the channel, not the chatbot
Instinct raised $1 billion in September 2026 at a $10 billion valuation. That number says something about how investors read the opportunity: the agent that owns your messaging thread owns the default route for a very large share of everyday intent. No app install, no habit to build, no cold start.
What I’d watch next is whether these systems converge on shared conventions for the hard parts—interruption, revision, permission, and graceful failure inside a 160-character window. Those are solvable problems, but they’re solved in the orchestration layer, not by swapping in a better model. The teams that treat the text thread as a genuine runtime rather than a thin client will be the ones whose agents feel less like a service you ping and more like a correspondent who remembers.
đź•’ Published: