Frontend

dang imagine your ai agent / copilot using this to make changes to your codebase then autonomously deploy to vercel or a github action 👀

Vercel TypeScript SDK

• Full type safety for accessing the Vercel REST API with Zod schemas
• New documentation for every function, argument, and type
• Better tree-shaking support with optional standalone functions

https://t.co/1rIv6nDbFm

𝘷𝘦𝘳𝘤𝘦𝘭.𝘥𝘦𝘱𝘭𝘰𝘺𝘮𝘦𝘯𝘵𝘴.𝘤𝘳𝘦𝘢𝘵𝘦𝘋𝘦𝘱𝘭𝘰𝘺𝘮𝘦𝘯𝘵()

You can now integrate and automate the @vercel platform with a simple TS SDK. This will be great for AI agents that deploy on users' behalves.

Guillermo Rauch

dang imagine your ai agent / copilot using this to make changes to your codebase then autonomously deploy to vercel or a github action 👀

https://x.com/rauchg/status/1866209983588900917

I was actually thinking about this concept last night before I went to sleep.

It would be really cool if talking to an agent created memories that update the system prompt instead of just being stored in some database.

Just as humans learn a lot from their environment, so too could the agent.

But in this case, the learnings would actually impact its *personality* rather than just being a memory.

ahh that'd be cool! there's this tool i've seen called braintrust that can run llm evals like tweaking your prompt based on responses/response quality, but i wonder if that's more meant for an agent that's single purposed(like it browses the web and you wanna make sure the llm doesn't hallucinate)

This is cool, thanks. gonna look into it. 

The thing holding me back from the idea I proposed above is that I don't yet have a solution for how to distinguish what should update the system prompt by how much.

On the surface, someone could just go back and forth with the bot and feed it a bunch of garbage, effectively hijacking the personality of the bot.

So there needs to be some decision-making framework ahead of time such that the bot retains its core personality while evolving given its interactions.

Open to ideas if you have any!

ahh that'd be cool! there's this tool i've seen called braintrust that can run llm evals like tweaking your prompt based on responses/response quality, but i wonder if that's more meant for an agent that's single purposed(like it browses the web and you wanna make sure the llm doesn't hallucinate)

https://www.braintrust.dev