Wabi rebuilds itself as a chat-first AI agent that builds apps on demand

Wabi, the AI startup that turned text prompts into working apps, has rebuilt itself around a chat-first agent that generates interfaces on demand, founder Eugenia Kuyda said this week as the company rolled out Wabi 2.0 by invite code.

The pivot lands the same week OpenAI and Meta pushed their own takes on agent-style assistants. It also reframes Wabi away from the crowded vibe-coding market and toward a more competitive fight for the default consumer AI surface.

What changed at Wabi?

Wabi launched as a tool for non-developers to build apps through plain-language prompts. The company now describes Wabi 2.0 as a messenger that still builds apps, but renders them inline inside a chat thread instead of as the product itself.

Kuyda, who previously founded AI companion startup Replika, framed the shift in plain terms. Wabi 2.0 is, in her words, "a personal agent that does stuff for you and builds the interface you need in the moment."

The reframing moves Wabi's competition. The new set of rivals includes AI agents from Meta — the recently introduced Muse and Instinct lines — plus any consumer chatbot that fuses conversation with task execution. The previous Wabi pitch placed Wabi next to other vibe-coding tools. That market remains crowded and capital-intensive.

Why is Kuyda betting against chat-only agents?

In a thread on X, Kuyda argued that pure chat interfaces break down under real-world use. "The more you use a chat-only agent, the worse the experience gets," she wrote. "Your history and ongoing tasks end up buried in one endless scroll."

She pushed the point further. "People don't just want to type-type-type; They want to tap-tap-tap, scroll, and look. They love what agents can do, but they also like software," Kuyda wrote. Her conclusion: "The most powerful agent will be the one that combines both."

That thesis tracks a broader consumer pattern. Voice assistants, search boxes, and messaging apps all won mass adoption by mimicking behaviors people already understood. Kuyda is betting that on-demand UIs, generated when the agent decides the user needs a tracker, log, or form, will feel more native than another chat scroll.

The argument is also a defensive one. If the entire category converges on chat, every product becomes a skin on top of one of three or four foundation-model providers. A company that owns a distinctive interaction model has a better chance of holding pricing power and user attention.

How does Wabi 2.0 actually work?

The core mechanic stays familiar. A user asks the agent to perform a task. The agent either responds in chat or generates a small, task-specific interface to complete the request.

Kuyda used a calorie tracker, a weightlifting log, and a weight tracker as examples in a thread about health. The same logic applies to family workflows. In another thread she described a kids' activities calendar, a language-learning app, and a local-events tracker for children.

Each interface lives inside the messenger rather than as a standalone app download. The user never has to leave the chat to enter a weight or check a soccer practice time.

That choice changes what Wabi is selling. The unit of value shifts from "app you built" to "interaction the agent chose to render." A user who builds one app and forgets about it churns. A user who messages an agent daily, and occasionally taps into a calorie tracker or schedule the agent surfaced, does not.

Kuyda laid out the longer-term vision in another X post. "So far, personal agents have been one person, one chat, a stream of one-off tasks," she wrote. "We think the future is your agent, your apps, and your people in one place, running the parts of life that matter most, together."

The "your people" line hints at a social layer Wabi has not yet detailed publicly. A messenger product without any way to share or collaborate would be an odd messenger product to ship.

Where does Wabi fit against OpenAI and Meta?

The timing is not accidental. The same week Wabi announced its pivot, OpenAI used Dev Day to update ChatGPT's plugin system. Apps inside ChatGPT now get a more visible spot in the sidebar, and developers can build interactive experiences that run inside the chat surface.

That update points the industry at the same conclusion Kuyda reached independently. The chat window is becoming the operating system. Plugins, on-demand UIs, and inline apps all compete for the same real estate inside the message stream.

Meta's Muse and Instinct agents push the same direction from the social-graph side. Both products aim to blend conversation with task execution, and Meta's distribution gives them an audience Wabi cannot match on reach alone.

Wabi's counter is product specificity. Where ChatGPT serves as a launchpad for third-party apps, Wabi plans to generate them on the fly inside the same thread. The trade-off is reach versus perceived responsiveness, and the bet is that consumers will value the responsiveness more.

OpenAI's plugin update also raises the stakes for any independent agent product. The more surfaces a foundation-model lab owns — chat, plugins, apps, voice — the harder it becomes for a startup to justify a separate app at all. Wabi's response is to argue that the chat-only model is itself flawed.

What's next for Wabi?

Wabi 2.0 ships in invite-only mode. Codes are being distributed on Kuyda's X account, the same channel she used to explain the product rationale. The company has not published a timeline for general availability.

Wabi has not disclosed user counts, revenue, or any new funding tied to this release. The next month of usage data will determine whether the messenger framing earns a wider rollout or retreats back to the prompt-to-app pitch that started the company.

The bigger question is whether the agent-plus-UI thesis holds outside a founder thread. If consumers do want tap and scroll inside their agents, the category will reward whichever vendor gets the interaction model right first. If they keep defaulting to typed chat with ChatGPT, Wabi's pivot becomes another cautionary tale in a year full of AI reshuffles. For now, Wabi is asking early users to test that bet, one invite code at a time.