Temporal Makes Multi-Framework AI Integrations GA; Workflow Replay Is Separate From Preview Streams

Temporal announced on October 30 that integrations with several AI libraries and agent frameworks are generally available. This November 1 technical issue retains October 30 as the original publication date. The GA designation describes the integrations identified in the announcement. It must not be extended to every adjacent capability: Workflow Streams for token streaming are separately described as Public Preview for Python and TypeScript.

The integration matrix is precise. Google ADK is listed for Python, Go and TypeScript. OpenAI Agents SDK is listed for Python, which was already GA, and TypeScript. AWS Strands covers Python and TypeScript; LangGraph covers Python; LangSmith covers Python and TypeScript; Vercel AI SDK covers TypeScript; and Gemini SDK covers Python. A team should select the package and guide for the language it actually runs. A general statement about “AI framework support” does not make all framework-language combinations interchangeable or grant an undocumented feature to a listed package.

The central design boundary is Temporal's separation of coordination from external work. A Workflow contains long-lived coordination: what should happen and when execution may wait. An Activity performs a model call, API request, database operation or other side effect. The integrations map framework behavior across that boundary, so an existing agent need not be rewritten around a completely different framework. In the Google ADK Python example, `TemporalModel` replaces a model name and makes each model call an Activity. Registering `GoogleAdkPlugin` with the Temporal Client and Worker makes the agent loop safe to replay. The OpenAI Agents SDK, Strands and LangGraph have their own integration points, so the ADK snippet is an illustration, not a universal installation recipe.

Recovery depends on Event History. If a Worker restarts or a deployment interrupts execution, Temporal reconstructs the Workflow from recorded progress. On replay, a previously completed Activity result is reused rather than making the same model or tool request again. That behavior matters for agents whose calls can spend money or change external state. It is conditional on placing side effects in Activities and using the integration as designed. A model call performed directly in replayed coordination code would not acquire the same guarantee simply because the application imported a Temporal package.

The announcement also discusses retry ownership. Integrations disable, or ask developers to disable, framework-level retry loops so that Temporal can manage retries and timeouts in one place. Without that discipline, two retry policies can multiply attempts and make failure analysis difficult. Client objects and credentials stay on the Worker, not serialized into Workflow history. Trace context crosses process boundaries, and replay should not export the same spans again. These are operational details behind the phrase “durable execution”; omitting them could turn a recovery demonstration into an unsafe production assumption.

Agents may wait for a human decision or other external input without polling continuously or keeping a Worker process occupied during the wait. Where supported, streaming tokens can pass through Workflow Streams, but the post calls that feature Public Preview rather than GA. The distinction is useful for release planning: an organization can adopt a GA integration while treating streaming as a separate compatibility and risk decision. One Workflow can also coordinate work from different frameworks, such as a LangGraph graph followed by a Google ADK agent and a human response, according to Temporal's example.

The released packages are hosted in `temporalio/ai-integrations`, with separate packages by framework and language. Temporal gives examples for Python, TypeScript and Go and directs upgraders to release notes. Existing Workflows deserve special care because package and import changes, and sometimes Worker Versioning, may be needed to let older executions finish under earlier code. A sensible evaluation is to choose the exact package, inspect its runnable sample, run a controlled interruption test, and check the Event History for duplicate external calls before changing a production workflow.

This publication has not measured a recovery success rate or audited every integration implementation. The grounded claim is that Temporal announced these named integrations as GA and explained a common execution model built on Workflows and Activities. The design judgment for adopters is to test the specific framework and language combination, preserve the replay boundary, and keep the preview streaming feature separate from the GA integration claim.