TanStack AI 1.0: stable application APIs, an experimental agent runtime

An internal AI application needs more than a chat box. It must know which customer owns a conversation, what a person approved and which actions already happened. That is how I would evaluate TanStack AI 1.0.

The October 23, 2026 announcement makes the core APIs stable while introducing an experimental agent harness for longer-running work. The new harness is not covered by the 1.0 stability promise. This topic came from supplementary primary-source research, not a release inferred from a Czech video.

Where I would use it

A custom support interface or internal CRM is a good candidate when the team needs control over both the interface and agent tools. A working n8n or Make process does not need rewriting just because a library reaches 1.0.

My first pilot would use twenty synthetic support tickets. The agent suggests a category, retrieves supporting material and drafts a reply. A reviewer sees both the source and the proposed change. Sending happens in a separate step with a recorded outcome. This is a proposed evaluation, not our own TanStack benchmark.

Migration includes stored data

The migration guide switches conversation identity to threadId, changes approvals to interrupts and requires handling older saved data. Client and server need to move together.

I would back up the data, test a copy of real conversations and write a rollback procedure. Passing TypeScript is only one check. Load an old chat, resume a pending approval, display an attachment and reconnect after a page reload. Otherwise, the new build may work only for new users.

Surviving a restart

The harness documentation distinguishes in-memory storage from a durable session log. The example's memoryPersistence() does not survive a restart. The deployment documentation also describes recovery limitations.

In the pilot, I would stop the process immediately before a CRM write and again immediately after it. Check for duplicate records, lost approvals and a traceable run owner. An operation identifier and deduplication belong in the target application, not merely in the model's instructions.

For a CTO, this warrants a controlled evaluation. A stable library helps integration; application-level tests must establish whether the entire process is dependable.