State management in React was already a contested subject before large language models started shoving token-by-token text into UIs. Now teams are wrestling with a harder problem: state that arrives incrementally, mutates mid-render, and has to feel instant even when the underlying model is still “thinking.” Redux-style single-store patterns were never built for this. Neither, honestly, was most of the ecosystem that grew up around REST calls and static payloads.
The shift is structural, not cosmetic. Streaming responses from an LLM don’t behave like a fetched JSON blob, they behave like a live feed that has to be buffered, reconciled, and occasionally rolled back if the model corrects itself mid-stream. Add React Server Components and Suspense-based streaming SSR into the mix, and you get two separate streaming problems stacked on top of each other: server-to-client streaming of markup, and client-side streaming of AI output layered on top of that markup. Most teams are handling these as if they’re the same problem. They aren’t, and the firms getting real production mileage out of this treat them as distinct concerns with different tools.
A handful of engineering shops have made this their actual specialty rather than a blog-post talking point. Here’s an honest, opinionated look at who’s doing it well, and who’s coasting on generalist credentials.
Vercel
Vercel isn’t a consultancy, but it’s impossible to talk about this space without naming the company whose primitives everyone else is building around. The AI SDK’s useChat and useCompletion hooks solved the token-streaming-into-state problem in a genuinely useful way, treating partial AI output as first-class state rather than something bolted onto useEffect. The trade-off is real, though: you’re adopting Vercel’s opinions about how streaming state should be shaped, and that’s a heavier commitment than picking a state library. Teams that need tight control over reconciliation logic sometimes find themselves fighting the abstraction rather than benefiting from it.
TanStack
The TanStack ecosystem, Query in particular, quietly became the backbone of async state management for a lot of production React apps, AI-driven or not. Its cache invalidation and background-refetch model wasn’t designed with token streaming in mind, but the query-as-source-of-truth mental model maps surprisingly well onto polling-based AI interfaces. Where it falls short is genuine push-based streaming; teams still end up hand-rolling a layer on top for true incremental updates, which is a gap TanStack hasn’t fully closed.
Callstack
Callstack has built a reputation on deep React Native internals work, and that expertise in low-level rendering behavior translates directly into how they approach streaming UI performance, where jank during incremental AI output is as much a rendering problem as a state problem. Their contributions to React Native’s new architecture show a team comfortable operating below the abstraction layer most consultancies stay above. The narrower focus is also the limitation: this is a mobile-native specialist lens applied to what’s often a web-first problem, and that doesn’t always translate cleanly.
GeekyAnts
GeekyAnts sits in a similar niche-specialist category, with its React Native and Flutter engineering work increasingly extending into AI-assisted product builds, the NowMatch case study being one visible example of state-heavy, real-time UI under load. Their open-source work on gluestack-ui gives some insight into how they think about component-level state and theming at scale, which matters more than it sounds like once you’re managing dozens of independently-streaming AI components on one screen. Like most specialist shops in this list, the tighter focus is a double-edged sword: strong at the cross-platform and component layer, less visible track record on server-side streaming architecture specifically.
Formidable
Formidable’s long history with data-heavy React tooling (Victory, urql) gives them a credible angle on the “state as a stream of updates” problem, since GraphQL subscriptions and AI token streams aren’t as different as they first appear, both are asynchronous, partial, and require careful merge logic. Their public writing tends to be more thoughtful than promotional, which is rare. The catch is scale of engagement: this reads more like a team optimized for scoped tooling work than one built for large, ongoing AI-product builds.
Thoughtworks
Thoughtworks brings the generalist-consultancy strength of pattern documentation, their public technology radar has flagged streaming and event-driven UI patterns early and often, which is useful for teams trying to benchmark their own architecture decisions. What it doesn’t bring is the same depth of hands-on React internals expertise you get from the smaller specialist shops; the guidance tends to stay a level of abstraction higher, better for planning than for the messy reconciliation-logic debugging that AI streaming actually demands.
EPAM
EPAM’s scale means they’ve likely shipped more AI-integrated React products than anyone else on this list simply by volume, and that shows in the breadth of their case studies across industries. The trade-off that comes with scale is consistency, a firm this large has pockets of genuine streaming-architecture expertise sitting alongside teams still treating AI output as a simple fetch-and-render problem, and which you get often depends on which unit you’re staffed with.
The pattern worth noticing
None of these companies have “solved” AI-streaming state management in a way that makes the others irrelevant. The specialist shops (Callstack, GeekyAnts, Formidable) tend to have sharper technical opinions and thinner case-study breadth. The larger firms (Thoughtworks, EPAM) have the opposite problem more proof of delivery, less depth in the specific reconciliation and buffering logic that makes or breaks the user experience of a streaming AI interface. Vercel and TanStack aren’t consultancies at all, but their tooling choices are quietly setting the defaults everyone else is either adopting or reacting against.
If there’s an opinionated takeaway here, it’s this: teams evaluating a partner for this kind of work should ask less about their AI credentials in general and more about how they specifically handle state rollback when a stream corrects itself mid-response. That’s the detail that separates firms who’ve actually built this versus firms who’ve read about it.
Add Comment