Lars Grammel

Humans have been writing software for about 75 years. AIs will write software for hundreds and thousands of years.
Eventually everything - hardware, OS, libraries, APIS, services, applications - will be designed to facilitate AI software development.

Harness engineering: leveraging Codex in an agent-first world By Ryan Lopopolo, Member of the Technical Staff

one of the best articles on the future of agentic coding that i've read so far: openai.com/index/harnes...

Been thinking a lot about git lately and how incredibly well designed it is.

People are scared of how AI will impact their lives. Shooting the messenger is their reaction.

bc you need many users to have a data flywheel that covers the long tail of user requests (for your training data), and free models do not have that. they are worse than they look at first glance (due to benchmaxxing)

future software will be developed in dark software factories by a variety of ai agents operating on a shared code repository infrastructure optimized for those agents. humans won't look at the code any more, similar to how we do not look at assembly today.

in a future where ai agents write all the code, libraries and frameworks are somewhat dead bc they include overhead unnecessary for the specific use cases of the software and represent a supply chain risk

YouTube video by Callstack AI SDK 5: What’s New and What’s Next by Lars Grammel | AI Meetup in Wrocław, September 2025

My talk about AI SDK 5 from the AI Meetup in Wrocław is on youtube: www.youtube.com/watch?v=xgR-...

prediction: as ai relies more and more on websearch, seo for ai will bring product placement and then it will be followed by ads for ai (content suggestions), esp. in free ai chatbot tiers. ai chatbots will become more like google search in some of those ways.

I find this research fascinating because it means that there is a path towards smaller models that are strong on tool use and reasoning and that use external knowledge sources (through tools such as web search) instead of baked in knowledge.

openai.com/index/why-la...

GPT-5 Thinking in ChatGPT (aka Research Goblin) is shockingly good at search “Don’t use chatbots as search engines” was great advice for several years... until it wasn’t. I wrote about how good OpenAI’s o3 was at using its Bing-backed search tool back …

The previously sensible advice to never use ChatGPT for search needs to be rethought - GPT-5 in thinking mode is shockingly good at running searches now simonwillison.net/2025/Sep/6/r...

At React Universe Conference Wroclaw today - would love to chat with AI SDK users.

If you are around and want to give feedback or learn more about the AI SDK send me a DM.

Finally - almost 5 months of work, our biggest release so far.

Completely revamped, type-safe AI chat is a game changer imo. I don't know any other framework that has full stack support for this. Hope it will enable you to build the next generation of AI applications.

i recommend cursor. tried switching back to copilot at some point to give it another shot and switched back to cursor almost immediately.

is that with the new chattransport?

You can try AI SDK 5 now:

AI SDK 5 prototype: custom UI data parts

You can define custom ui data parts and stream updates to them from the server:

AI SDK 5 preview: ChatStore & ChatTransport

ChatStore synchronizes chat write operations and manages streaming state. You can use it directly or through framework integrations like useChat.

ChatTransport makes backend integrations more flexible, allowing for client-only usage.

AI SDK 5 preview: UI message persistance

We recommend storing UI messages, not model messages, if your application has a useChat component.

This ensures that the UI state can be correctly restored and helps integrating backends other than streamText (e.g. Langchain).

For agent to agent communication you have to set up something custom - focus for AI SDK 5 is to improve assistant to user messages in the UI

AI SDK 5 preview: UI message metadata

UI messages in AI SDK 5 will have a generic metadata property (instead of specific properties like createdAt).

This lets you send and show the message metadata that's important in your application:

It's important to distinguish UI messages from model messages.

Having a clear separation makes it easier to set up (application specific) persistence for your UI messages and additional context. It also helps putting prompt optimizations in place and improving results.

The model prompt is often e.g. enriched and compressed.

E.g. short-term memory compresses, long-term memory and RAG enriches, system messages add additional information. You could have different model message prompts depending on the context.

UI messages contain additional information such as custom app data and metadata. Tools calls might be omitted. etc.

UI messages != model messages

What you display to the user (ui messages) is different from what you want to send to the LLM (model messages).

UI messages contain additional information such as custom app data and metadata. Tools calls might be omitted. etc.

In AI SDK 5, UI messages will be composed of parts.

This is my current vision of how UI messages could look like - feedback welcome.

Starting to be excited about AI SDK 5

Our planned timeline:

- alpha 1st half of May
- beta 2nd half of May
- GA in June

You can improve agent quality (i.e. success within a certain number of steps) with better prompts, tools, and other scaffolding - but you always need to account for the runaway run problem by limiting steps somehow (e.g. using a budget like tokens, steps, or money spent)

AI agents vs workflow runs & run termination:

Workflows have a deterministic number of steps after which they will finish with with a solution (of varying quality).

Agents run in a loop and it is unclear when and if they will finish with a solution.

one fundamental issue w/ ai agents:

the problem of deciding when to stop an agent run is a variant of the halting problem

there will always be agent runs for which, given their current state and history, it cannot be known whether they will terminate with an answer or not