man AI SDK v5 is so good, feels bad having to build agents without it whats the equivalent or other languages (python etc)

typed custom data and tool calls on the client are the main dx innovation in ai sdk 5

Our Next.js, Nuxt, and Svelte quickstart guides have been updated to AI SDK 5 Beta. Get started today.

"working as expected"

I would also add agentic control with prepareStep and stopWhen What kind of templates/videos would you like to see?

sometimes you need to stream ephemeral data to the client as part of an assistant response you can now mark data parts as transient to prevent them from becoming part of the message parts, and use the onData callback to react to them streaming in

Chat SDK is now using AI SDK 5 Beta.

My workshop from Vercel ship is now live! Build a coding agent like Claude Code with AI SDK 5, AI Gateway, Vercel Sandbox, and Vercel Functions in 40 minutes

We added an experimental Agent abstraction to the AI SDK. It has the same functionality as streamText/generateText (which are used by it), differently packaged for those who prefer an explicit agent object.

Lars Grammel Went a bit overboard and did you a personal YouTube video, enjoy gist.github.com/mattpocock/ff8…

The AI SDK v5 is really, really good - Cuts out a lot of crap - Adds flexibility - Tons more type-safety - More LLM features

ai sdk 5 agent with tools

Can't stop cooking AI SDK v5 beta has got my juices flowing

you can now consume ui message streams without useChat, e.g. for terminal UIs or custom frontend hooks

Excited to host Lars Grammel, core maintainer of the AI SDK at Vercel, for the first installation of Langfuse Context at our langfuse.com Berlin office. Join us to learn more about his journey, take on LLM/agent libraries, and insights on what's next! sign up below

this is great for prompt compression, adding caching options, and injecting memory

long-running tasks with stopWhen (aka maxSteps) can very quickly clog up your context window with prepareStep, you now have complete control over everything that is being sent to the model at the beginning of EACH step