The dominant story in agent design right now says the interface should get richer. Bigger canvases, browser panes, inline tool output, streaming traces you can inspect. The text-message agents that showed up over the past year argue the opposite, and I think they’re closer to right. Constraint is not a limitation these systems tolerate. It’s the thing doing the work.
TechCrunch compiled a list of the most notable agents that now live inside your messages, spanning general assistants, family-oriented tools, and professional task handlers. Reading it as an architecture document rather than a product roundup is more interesting than it sounds.
What a message thread actually forces you to build
Strip an agent down to a plain text channel and several design decisions stop being optional.
- Asynchrony becomes the default. A chat window implies the user is watching. A text thread implies they are not. That single assumption changes the control loop: the agent must be able to start a task, lose its reader, and resume coherently when a reply arrives hours later.
- Output has a hard budget. There is no collapsible panel to hide reasoning in. Anything the model emits, the user reads. Verbosity, which most chat interfaces quietly subsidize, becomes an immediate cost.
- Identity is bound to a phone number. No session tokens, no login screen, no per-tab state. The thread is the session, and it persists across months.
- Interruption is a first-class event. Users send three messages in a row, change their minds mid-task, and reply to something from last Tuesday. The agent needs conflict resolution over its own task queue, not just turn-taking.
None of that is new theory. What’s new is that a delivery channel enforces it. You cannot ship a sprawling, chatty, stateless agent into SMS and have it feel usable. The medium does the pruning your product review should have done.
Platform approval as an architectural signal
Poke became notable in June 2026 when Apple approved it for the Apple Messages for Business platform. Apple’s approval matters less as a distribution milestone than as a statement about what the platform owners consider acceptable behavior inside a messaging channel they control.
Messaging rails are not open web. They come with deliverability rules, rate expectations, and consequences for abuse that land on the sender, not the user. An agent operating there inherits a governance model it did not design. For anyone building in this space, that’s the real cost of entry: your retry logic, your proactive notifications, your error handling all become compliance surfaces.
Three different bets on the same channel
The agents on the list are not competing on the same axis, which is easy to miss.
Trust-first
Ollie is a family-focused assistant with SOC 2 compliance. Positioning a consumer family tool around an audit standard usually reserved for enterprise procurement is a deliberate choice. When an agent has standing access to a thread containing a household’s logistics, the security posture is not a feature bullet. It’s the product.
Task-first
Town offers a professional assistant handling tasks across various platforms. This is the hardest engineering problem of the three, because cross-platform execution means the agent’s tool layer is wider than its interface. A one-line text reply may sit on top of half a dozen authenticated integrations, each with its own failure mode that has to be summarized honestly in a sentence.
Capability-first
Genspark offers unlimited access to multiple AI models for chat and image generation in 2026. Multi-model routing inside a single thread is an orchestration bet: the agent decides which backend serves each request, and the user never sees the handoff. Done well, model choice becomes invisible infrastructure rather than a setting.
Worth separating from all of this: Cognigy leads enterprise voice with a native Voice Gateway, and sits among the tier-1 packaged enterprise support vendors alongside Sierra, Decagon, and Salesforce Agentforce. CloudTalk’s AI VoiceAgents work the phone line, handling inbound calls, order and ticket status, appointment booking, FAQs, and routing. Voice and text share a surface-level similarity — both are conversational, both are low-bandwidth — but voice adds real-time latency constraints that text does not have. Treating them as one category produces bad engineering.
What I’d watch next
Pricing remains unclear across most of these products, which makes unit-economics analysis premature. The more tractable question is memory. A thread that lives for a year accumulates state no context window holds. Whoever solves long-horizon memory inside a channel with no UI for managing it will have built something the richer interfaces can then copy.
The reduced surface was never the compromise. It was the specification.
🕒 Published: