A stability promise, stated plainly
Most 1.0 announcements bury the commitment. TanStack puts it near the top: 1.0 is “our formal commitment to the APIs you build on”, everything in the 76 packages is stable and ready for production, and, unusually, the release adds no new concepts. Every API that 0.x marked as deprecated was removed, and every replacement already shipped during 0.x. The features that arrived since the August release candidate, more than 100 merged pull requests by the team's count, are refinements of a toolkit that already covered chat, tools, media generation, MCP, persistence, sandboxes and agents.
That restraint is the strongest part of the release. A team that kept up with the 0.x deprecation warnings is, in TanStack's words, most of the way there. The growth since the alpha is still substantial: four provider adapters became 23, and sandbox providers went from none to 11. On testing, TanStack reports that its deterministic end-to-end test suite grew from 265 tests across 10 providers at Beta to 943 tests across 17 provider adapters on the 1.0 branch. The toolkit lists 23 provider adapters overall, so that test figure should not be read as equivalent coverage of all 23. Those are the project's own CI figures, not an independent audit — but they are specific enough to be checked against the repository, which is more than most toolkit launches offer.
The exception is the headline feature
The post then spends its longest section on the one thing the stability commitment does not cover. The agent harness turns a single chat() request into a session that stays open: durable session logs that let a turn cut by a crash or a deploy continue on the next host, approvals, questions and sign-ins that wait in the log until someone answers, typed agents run in parallel or in series with budgets and timeouts, plugins, background workers, and a set of first-party plugins that assemble a terminal coding agent able to hand work to Claude Code, Codex, Grok Build or any ACP agent, each in its own sandbox.
It is, plainly, the part of TanStack AI aimed at where agent development is heading. It is also shipped as @tanstack/ai-harness and @tanstack/ai-harness-cli at version 0.x, under a section headed “Experimental, on purpose”. The harness APIs can change in a minor release, and the harness entry points inside other packages — the post names @tanstack/ai-skills/harness and @tanstack/ai-acp/agent — are experimental too. TanStack's stated reason is that it wants the harness in developers' hands before locking it, and the section closes with the boundary in one line: “Everything else in this post is 1.0.”
One thing the post never settles is the arithmetic. It says all 76 packages go stable together, and it calls the harness “the one exception”, but it publishes no package inventory — so whether the two 0.x harness packages are counted inside the 76 or sit alongside them is not defined by the announcement itself, and this article does not resolve it either. What the post defines precisely is the boundary that matters to users: harness APIs sit outside the stability promise, wherever they are exposed. A reader who stops at “everything is stable” will miss that the newest capability carries the opposite guarantee. The accurate summary is two releases in one post: a stable toolkit, and an experimental API surface attached to it — useful for evaluation and early implementations, but subject to breaking changes before it reaches a stable release.
What upgrading actually asks
The upgrade notes deserve the same close read. Every @tanstack/ai* package must move together, because a 1.0 client needs a 1.0 server. Three packages — the older React, Solid and Vue UI packages, @tanstack/ai-react-ui, @tanstack/ai-solid-ui and @tanstack/ai-vue-ui — get no 1.0 at all; their users are told to remove them and rebuild the interface with createChatHook from the corresponding main package. And threadId, the identifier for a conversation, is now required on both sides, with neither side generating one for you. The codemod handles renames such as body becoming forwardedProps and conversationId becoming threadId, but it cannot add a threadId, because it does not know where your conversation IDs come from. Teams holding saved chats, runs or files from 0.x have separate data steps to complete before deploying.
None of that is unreasonable, and the guide ends with a done check so teams know when they are finished. It does mean 1.0 is not a version bump you apply casually on a Friday. For new projects, the calculation is simpler: the stable core is a credible way to avoid committing to one model provider, one UI framework or one sandbox at the start. Just budget for harness API changes, and keep anything you build on the harness isolated enough to change when it does.
Sources
- [1] TanStack, “TanStack AI 1.0: Still Your AI, Still Your Way”, TanStack Blog, 23 October 2026 — primary source, read in full first-hand for this article: “All 76 packages go stable together” and “Everything in the 76 packages is stable and ready for production”; the harness exception (“The new agent harness is the one exception. Its APIs are experimental.”; @tanstack/ai-harness and @tanstack/ai-harness-cli ship as 0.x; harness entry points in other packages, naming @tanstack/ai-skills/harness and @tanstack/ai-acp/agent, are experimental too; “Everything else in this post is 1.0.”); the comparison table (23 provider adapters plus openaiCompatible; UI integrations React, Vue, Solid, Svelte, Preact, Angular, Octane, Remix and vanilla JS; 11 sandbox providers); the test figures (265 across 10 providers at Beta, 943 across 17 provider adapters on the 1.0 branch; more than 1,800 docs samples type-checked in CI); more than 100 feature pull requests since the RC; 143 contributors and a core team of three maintainers; and all upgrade conditions, including the three UI packages with no 1.0, the required threadId and the codemod's limits. The post publishes no package inventory defining whether the two 0.x harness packages are counted within the 76.Read source
- [2] TanStack AI documentation index, TanStack (cross-check of the post's date, 23 October 2026, and its summary line: “76 packages go stable together, deprecated APIs are gone, the docs are rewritten, and an experimental agent harness lands”)Read source
- [3] AI Agent News Today — October 23, 2026, AIAgentStore (secondary roundup carrying TanStack AI 1.0 as a new 23 October item, with the 76-package and 23-adapter figures; used only as a freshness cross-check, not as a source for any claim in the article)Read source